FILE kostic-project-analysis.md

DATE 2026 · 05 · 26

STATUS active

IDENTITY work

Columbia EECS · Summer 2026 · Interview prep · Vol. 1 of 2

Kostic project
deep-dive — all twelve subprojects, analyzed

A project-by-project read of AIDL's open summer slots, scored against a real record in hardware-systems and applied AI. Where the fit is honest, where the crowd will cluster, and what to walk into the Zoom with.

Subject

Prof. Zoran Kostic
AIDL Lab · Columbia EE

Interview

Wed 2026-05-27
3:00pm EDT · Zoom

Projects analyzed

11 subprojects
+ 1 standalone (AI on the EDGE)

Primary target

AI on the EDGE
Secondary: Distillation SAM→YOLO

Companion to kostic-research-projects.md

DOC · KP · A1 · 2026-05

00 How to read this

A six-section read, ranked by fit against a real record.

Fit is scored against the actual portfolio — a 2-way out-of-order RISC-V core with a 27.5% geomean cycle cut, the DE1-SoC VGA pipeline plus Linux kernel driver, ESP32 / UAV work under real-time constraints, Palimpsest NYC's agentic loop at 75.5% citation correctness, and the Gemma rank-1 ablation / distillation study. Where a project doesn't touch those, the score says so.

Audienceself, pre-interview Length6 sections · ~22 pages Read ordercover → §1 → §2 → skim §3–4 → §5 last Companionkostic-research-projects.md Brief2026-05-27-kostic-seok-interview-prep.md

Sections

01 Who Kostic is, and how his lab actually worksProfile · lab thesis · the four-course training ladder · deliverables regime p. 04

02 Project-by-project analysis11 subprojects + AI on the EDGE, sorted into four tiers p. 06

03 Cross-project comparisonFit · popularity · HW-relevance · ownability · compounding p. 15

04 Career-development mappingWhich projects compound toward which careers p. 17

05 Interview playbook — tomorrow, Wed 5/27, 3:00pmLead line · two pitches · papers · questions · landmines p. 18

06 SourcesCourses · papers · lab links p. 20

30-second ranking

#	Subproject	Umbrella	Fit	Crowd	Why (one line)
1	AI on the EDGE	standalone	HIGH	HIGH	His core thesis = your exact pitch; HW-systems edge
2	Distillation SAM→YOLO	Environment	HIGH	MED-HIGH	Distillation + edge-deploy background, clean scope
3	Trajectory Prediction	Streetscapes	MED-HIGH	HIGH	DNN + optional LLM/VLM; modest HW lever
4	Real-Time Agentic Streetscapes	Streetscapes	MED-HIGH	VERY HIGH	Palimpsest agentic loop maps directly; trendy → crowded
5	Adaptive Data Collection	Streetscapes	MEDIUM	MEDIUM	ML-theory heavy; math-of-DL helps, not your core
6	3D Models for Streetscapes	Streetscapes	MEDIUM	MEDIUM	Real-time 3D is the hook; you lack 3D-vision background
7	Surgery — Agentic Vision	Surgery	MEDIUM	MEDIUM	Agentic + edge latency angle; domain unfamiliar
8	Surgery — Video World Models	Surgery	LOW-MED	MEDIUM	A4 edge-deploy is your only hook; diffusion-heavy
9	PCatcher (plastics)	Environment	LOW	LOW	Jetson/DeepStream is the only overlap
10	Boundless water-trash (UE5)	Environment	LOW	LOW	Pure UE5 synthetic-data; not your stack
11	Microplastics monitoring	Environment	LOW	LOW	Spectral imaging + environmental; weak match

11 distinct subprojects in the doc; "Surgery" umbrella adds the parent framing for 12 entries total. AI on the EDGE is a standalone project, not a subproject.

01 The lab

Who Kostic is, and how his lab actually works.

Twenty years in industry R&D before academia — Bell Labs, AT&T Research, Thomson, MathWorks, Broadcom. Roughly 36 patents. Deep roots in SoC and heterogeneous/parallel computing. The kind of advisor who notices when a candidate can speak about cycles and cache lines, not just losses.

TitleProfessor of Professional Practice, EE DirectorEE MS Program AffiliationData Science Institute PhDEE · University of Rochester Emailzk2172@columbia.edu Siteaidl.ee.columbia.edu ScholarTlPI8yIAAAAJ

The one-sentence lab thesis

Efficient applied AI with a focus on inference — making models run in real time on real hardware. That industry / systems lineage is why hardware depth is the rare, valuable half he doesn't see often.

The four-course training ladder

Every project draws on one or more. Course pages live on the AIDL site.

Course	Title	What it teaches	Backs
ECBM E4040	Neural Networks & Deep Learning	CNNs, RNN/LSTM, autoencoders/GANs, backprop; PyTorch / TF	Foundation for all projects
EECS E4750	Heterogeneous Computing for Signal & Data Processing	GPU / GPGPU, CUDA / OpenCL / Metal, memory hierarchy, tiling, profiling	Edge / GPU optimization, real-time CV
EECS E6691 /6791	Advanced Deep Learning	Transformers, ViT / SWIN / DETR, diffusion (DDPM / Stable Diffusion), CLIP, LLM agents, YOLOv3–v9, Mamba, GNNs	VLMs, diffusion surgery models, agentic vision, detection
EECS E6692	Deep Learning on the Edge ★	Jetson Nano / Orin, CUDA edge optimization, profiling, compression / quantization, edge↔cloud connectivity	AI on the EDGE, distillation, all edge-deploy tracks

Funding & testbeds

Name-drop sparingly to show you did homework — never as filler.

$26M

NSF CS3 Engineering Research Center · Smart Streetscapes

$3.8M

COSMOS³ supplement · PAWR wireless+edge testbed, West Harlem

$1.2M

NSF CPS · urban digital-twin

25k

A100 GPU-hours from NVIDIA grant + 2× Jetson AGX Orin

Also: NOAA microplastics grant · NIH R01 (speech/health).

The deliverables regime — know this going in

Cadence

Daily & weekly rhythm

Daily activity log
Daily GitHub commits
Weekly oral report
1–2 meetings/week (in-person full-time in summer)

End of term

Publish-and-ship

Conference-style engineering report
~24-slide deck
Oral presentation
Clean GitHub with READMEs

The caveat that sets your odds

His own doc: “It is very rare that a first-semester student is accepted,” and it's “desirable to take one of Prof. Kostic's courses before applying.” You're finishing your first semester with no Kostic course — a marginal case on paper. Your job in the interview is to make the hardware-systems differentiator outweigh that. (More in §5.)

02 Project-by-project analysis

Eleven subprojects, four tiers — scored against a real portfolio.

Each card carries the project's umbrella, fit, predicted crowd, the Kostic-backed papers / courses it draws on, a sketch of what a 12–14 week summer would look like, and an honest read on career compounding.

Tier 1 Strongest fit · his core thesis · hardware-systems differentiator carries weight

1

Standalone · primary target AI on the EDGE — Optimization of Inference for DL on Constrained GPU / IoT Devices ★

Fit · High Crowd · High Course · E6692 Devices · Jetson · Coral · NCS Up to 3 credits Limited paid

What it is

Optimize inference for deep AI on networked low-power portable devices. Partition LLMs and other DL models into distributed modules that fit limited memory and compute; share load with edge-cloud compute and data servers; measure and minimize latency; explore real-time constraints. Hands-on with NVIDIA Jetson Nano, Google Coral, Intel Neural Compute Stick. Apps in smart cities and healthcare.

Kostic-background link

Course: EECS E6692 Deep Learning on the Edge is literally this project in course form (Jetson labs, profiling, compression, edge↔cloud connectivity).
Paper · Distributed VLMs (PerCom 2025 Workshop): Vision encoder on edge, language generation on cloud, +33% throughput vs cloud-only. This is "partition models across edge and cloud."
Paper · SASS — Streetscape Application Services Stack: Multimodal sync (88% timing-error reduction), spatiotemporal fusion (>10% detection gain), distributed edge computing (>10× throughput).
Funding: The NVIDIA grant (Jetson AGX Orin + A100 hours) explicitly funds low-latency multi-sensor edge fusion without cloud dependency — this project's money.

What a summer looks like (≈12–14 weeks)

Wk 1–2

Set up a Jetson Orin; reproduce the Distributed VLMs split as a baseline; profile where latency / energy actually goes (profiling instinct shines).

Wk 3–6

Pick a model (an urban detector or a small VLM) and implement a partitioning scheme; instrument an end-to-end latency budget (SLA) across edge↔cloud.

Wk 7–10

Core contribution — e.g. a smarter split point under a latency SLA, or quantization / operator-scheduling on the edge half; measure latency / energy / accuracy Pareto.

Wk 11–14

Deploy on real COSMOS-style data, write the conference-style report, build the ~24 slides, present.

Fit vs. you — HIGH

His doc's wording ("partition LLMs into distributed modules," "share load with edge-cloud," "minimize latency," Jetson / Coral / NCS) is almost a transcript of your cold-email pitch. Assets line up cleanly:

DE1-SoC co-design + kernel driver → comfort at the HW / SW boundary and with memory-mapped data movement, which is exactly where edge-inference latency hides.
OoO RISC-V (27.5% cycle reduction) → you reason about pipelines, utilization, and where cycles go — rare among ML-only applicants.
Gemma distillation / ablation → credible on the model-compression half.
Coral / NCS → his doc names them, so the earlier mention is on-target (the lab's published work leans Jetson; frame Coral / NCS as "I've also worked with…").

Predicted popularity — HIGH

Edge ML is hot and the title is magnetic. But most applicants are ML-only; the differentiator is hardware-systems depth, which is your moat. You're competing on a crowded field but from an unusual angle.

Career value

Straight line into edge-ML / MLSys / inference-optimization roles (the most hardware-adjacent, fastest-growing ML niche) and into hardware-aware ML research. Compounds your existing "I build the silicon under the code" identity better than anything else on this list.

2

Environment umbrella · strong secondary Distillation for Low-Complexity AI Models ★

Fit · High Crowd · Med-High Courses · E6691 + E6692 Device · Jetson Scope · open ("TODO add details")

What it is

Distill the SAM model down to YOLO models for real-time object detection. Optimize compute and memory. Deploy on NVIDIA Jetson. The doc literally says "TODO add details" — so the scope is yours to help define, which is an opportunity, not a red flag.

Kostic-background link

Course: E6691 Advanced DL (distillation, model compression, YOLO evolution, SAM / foundation models) + E6692 (edge deployment).
Paper · Monocular 3D Tooltip Tracking (MDPI Electronics 2025): Uses exactly this recipe — Florence2 + SAM2 zero-shot, refined with YOLOv11 supervised training (84.5% → 91.2% Jaccard). The SAM→YOLO distillation pattern is already live in the lab.
Related: The Constellation / Boundless detection work is the data side of the same coin.

What a summer looks like

Wk 1–2

Stand up SAM / SAM2 as teacher and a YOLO student; reproduce a baseline detection pipeline on Jetson; measure the teacher's latency (slow) vs target.

Wk 3–6

Implement knowledge distillation — match the YOLO student to SAM's segmentation / detection on an urban or environmental dataset.

Wk 7–10

Push compute / memory optimization (quantization, pruning) to hit a real-time frame-rate budget on Orin; quantify accuracy lost vs latency gained.

Wk 11–14

Deploy, measure sustained FPS / energy, report + slides.

Fit vs. you — HIGH

This is the cleanest scoped, most "ownable in one summer" project for you. You already have distillation experience (the Gemma project), it ends in a hard latency number on real hardware (your comfort zone), and the open "TODO" scope lets you propose the exact contribution. Lower competition than AI on the EDGE because it sits under the less-glamorous "Environment" heading.

Predicted popularity & career value

Med-High. Distillation is trendy, but it's buried under "Environment," so casual applicants skip it. Good for you: real fit, thinner field. Same edge-ML / model-compression track as #1, with a concrete distillation artifact you can show. Strong portfolio piece for MLSys roles and PhD apps.

Tier 2 Solid fit · agentic loop transfers from Palimpsest · lighter HW lever

3

Smart Streetscapes Trajectory Tracking / Prediction

Fit · Med-High Crowd · High Data · COSMOS testbed Target · 20 FPS @ A100

What it is

Real vehicle / pedestrian trajectories from the COSMOS testbed. Replicate trajectory-prediction models for comparison; integrate traffic-light state + environmental constraints into the simulator; integrate LLM / VLM into high-level decision making.

Kostic-background link

Paper: Data-Driven Traffic Simulation for an Intersection in a Metropolis (CVPR 2024 Workshop) — two-stage trajectory generation (coarse sampling → TrajNet++ refinement), 0.36 FDE @ 20 FPS on A100.
Paper: Language-Guided Traffic Simulation and Rare Event Synthesis (CVPR 2026 Workshop) — natural-language control over scenarios, LLM-to-loss, rare-event synth.
Paper: Urban Transportation Digital Twin (IEEE T-ITS).
Course: E4040 (RNN / LSTM) + E6691 (transformers, LLM agents).

Fit · popularity · career

"Integrate LLM / VLM into high-level decision making" maps onto your Palimpsest agentic-loop work. Pure trajectory modeling isn't your background, but the LLM / agentic layer is, and there's a modest HW / real-time lever (20 FPS target). Less hardware than Tier 1. Crowd: high — self-driving + LLM is a magnet topic; much of the competition will be CV-native students with more trajectory experience than you. Bridges toward autonomous-systems / CV research and agentic-AI roles; good but less unique to your HW identity than Tier 1.

4

Smart Streetscapes Real-Time Agentic Models for Smart Streetscapes

Fit · Med-High Crowd · Very High Scale · 1000s of cameras · real-time Course · E6691

What it is

Implement agentic AI workflows for streetscapes (cf. Gemini "agentic vision"). Detect objects + weather in street scenes; find limitations in existing agentic workflows and improve them; run locally over 1000s of cameras in real time. Doc flags it as fast-moving / frontier.

Fit · popularity · career

Your Palimpsest project is a bounded two-tool agentic loop with honest pre-registered evaluation — directly transferable to "agent loop with bounded latency, find + fix limitations." The "real-time over 1000s of cameras" angle even invites your hardware instinct. Weaker on the CV-perception side. Crowd: very high — "agentic" is the single hottest keyword in the doc; this will draw the most applicants. Strong fit but the most crowded — your Palimpsest artifact is what separates you from the pack here. Agentic AI is a fast-rising hiring area; combined with the real-time / edge framing, it's a strong story for applied-AI and research roles.

5

Smart Streetscapes Adaptive Data Collection and Annotation

Fit · Medium Crowd · Medium Track · theory-leaning Paper · ICML 2025

What it is

Refine a theoretical framework for optimizing data collection for robust model training under multi-distribution scenarios; larger-scale experiments on classification, detection, segmentation, VLM. Uses COSMOS multi-view feeds. Skills: DNN training; LLM/VLM finetuning a plus; ML theory a plus.

Kostic-background link

Paper: Adaptive Data Collection for Robust Learning Across Multiple Distributions (ICML 2025) — UCB sampling + online gradient descent, minimax-regret guarantees. Course E4040 + ML-theory foundations.

Fit · popularity · career

Your Mathematics of Deep Learning (A) course and the Gemma rank-1 ablation show you can handle the theory / perturbation framing, but this is the least hardware project in the set and leans toward students who want an ML-theory trajectory. Honest read: doable, not your sweet spot. Theory-flavored projects self-select a smaller, more specialized applicant pool. Best for a theory / ML-research or PhD path; weaker for your hardware-systems identity.

Tier 3 Lighter fit · edge-deployment is the only hook · domain unfamiliar

6

Smart Streetscapes 3D Models for Dynamic Streetscapes

Fit · Medium Crowd · Medium Hook · real-time only Stack · SAM3D · Hunyuan3D · TRELLIS · Gaussian Splatting

What it is

Frontier 3D-vision models (SAM3D, Hunyuan3D, TRELLIS); combine 3D reconstruction with background in dynamic urban scenes; run in real time in complex streetscapes; tackle depth estimation / Gaussian splatting and inpainting (LAMA too low-quality, FLUX Kontext insufficient) to remove dynamic objects.

Fit · popularity · career

Only real hook is "run these in real time" — your optimization / HW instinct. You have no 3D-vision / Gaussian-splatting background, so you'd be climbing a learning curve on the core. Pitch only if you're excited to learn 3D vision. Cool but specialized; needs 3D-vision chops, which thins the field. 3D vision / graphics-ML — a different track from yours; broadens but doesn't compound.

7

Surgery Agentic Vision Models

Fit · Medium Crowd · Medium Domain · medical · Lenox Hill Hook · agentic + edge latency

What it is

Build an agentic vision system that constructs structured surgical context from live video (segmentation / detection tools) and reasons over it in real time, under latency constraints. Tasks: agent loop with bounded latency; perception tools (seg / detection / pose / depth); context construction + temporal memory; latency-constrained edge deployment.

Kostic-background link

Course E6691 (agents, foundation models); paper Monocular 3D Tooltip Tracking (Florence2 + SAM2 + YOLOv11) and the IPAL / Lenox Hill surgical collaboration.

Fit · popularity · career

Same Palimpsest-agentic + edge-latency transfer as #4, plus the explicit "deploy on edge under latency constraints" hook. But the surgical domain is unfamiliar and access likely depends on the Lenox Hill collaboration. Niche domain (medical) narrows the pool despite the agentic hotness. Respectable, domain-specific.

8

Surgery Video Models in Robot-Assisted Surgery (World Models)

Fit · Low-Med Crowd · Medium Track A3/A4 only · edge accel + Orin / Thor

What it is

Train / deploy video generation / world models predicting future surgical scenes in real time from live robotic camera input; low-latency on edge. Track A: video diffusion (Wan 2.2, LTX-Video 2), conditioned prediction, inference acceleration (reduced-step diffusion, LoRA, distill + compile), edge deploy on AGX Orin and AGX Thor (A4). Track B: kinematics + world models from depth / pose.

Fit · popularity · career

Hook is narrow but real: A3 (distill + compile, reduced-step diffusion) and A4 (edge deployment + latency / FPS measurement) sit right in your wheelhouse. But the bulk is video-diffusion training — not your background — and the surgical domain is unfamiliar. Pitch only the acceleration / deployment tracks if you raise this one. Diffusion + world models is trendy; surgical domain narrows it. Generative-video / world-models is a frontier area; the edge-acceleration slice connects to your strengths, the rest doesn't.

Tier 4 Honest weak fit · don't pitch these

Project 09 · Environment

PCatcher — AI-based plastics collection on water surface

Fit · Low Crowd · Low

Train object-detection models for water-surface obstacle detection (RGB and / or LiDAR); integrate NVIDIA DeepStream SDK on Jetson embedded devices.

Single overlap: Jetson / DeepStream embedded deployment, which your embedded background touches. Everything else (marine object detection, the physical collection device) is outside your record. Career value limited for your direction.

Project 10 · Environment

Boundless — Water-Surface Trash Detection via UE5

Fit · Low Crowd · Low

Use Unreal Engine 5 to generate a synthetic RGB dataset for water-surface trash detection (HQ RGB + depth + bounding boxes). One student.

Backing paper: Boundless (UE5 synthetic data, +7.8 mAP vs CARLA) + Constellation dataset (13K images). This is a UE5 / synthetic-data engineering role. No game-engine / 3D-asset pipeline experience in your record; nothing maps here. Off-track for career.

Project 11 · Environment

Microplastics — AI Real-Time Monitoring

Fit · Low Crowd · Low

Distinguish plastics from natural debris via spectral + color-camera imaging for selective waste collection; tested in Hudson / Bronx Rivers. NOAA-funded; Lamont / Climate-School collaboration.

Spectral imaging + environmental sensing; outside your background. Only thread is "real-time," which is thin here. Mention environmental interest only if asked; don't pitch.

03 Cross-project comparison

All eleven projects on five axes, side by side.

#	Subproject	Fit	Crowd	Paid / Credit	HW lever	1-summer ownable	Career compound
1	AI on the EDGE	High	High	limited paid / ≤3 cr	High	Yes	Best
2	Distillation SAM→YOLO	High	Med-High	limited paid / ≤3 cr	High	Yes — cleanest	High
3	Trajectory Prediction	Med-High	High	≤3 cr	Low-Med	Yes	Med
4	Agentic Streetscapes	Med-High	Very High	≤3 cr	Med	Partly (frontier)	High
5	Adaptive Data Collection	Med	Med	≤3 cr	Low	Partly (theory)	Med (theory track)
6	3D Models	Med	Med	≤3 cr	Med (real-time)	Risky	Med (new track)
7	Surgery — Agentic Vision	Med	Med	≤3 cr	Med	Partly	Med (medical)
8	Surgery — Video World Models	Low-Med	Med	≤3 cr	Med (A3/A4)	No	Med
9	PCatcher	Low	Low	≤3 cr	Low-Med	—	Low
10	Boundless	Low	Low	≤3 cr	Low	—	Low
11	Microplastics	Low	Low	≤3 cr	Low	—	Low

Where the crowd will cluster

Popularity synthesis

Agentic + edge + LLM/VLM draws everyone

The applicant crowd will cluster on the agentic + edge + LLM/VLM labels: #4 (Agentic Streetscapes) draws the most, then #1 (AI on the EDGE) and #3 (Trajectory). Environment (#9–11) and the theory project (#5) draw the fewest. Surgery (#7–8) is gated by domain interest and collaborator access.

What that means for you

Be the rare hardware-systems person on a crowded ML project

Your edge isn't being one more ML applicant on the hot projects — it's being the rare hardware-systems person on a project where everyone else is ML-only. That argues for #1 (AI on the EDGE) as primary and #2 (Distillation) as the high-fit, lower-competition fallback. Let Kostic steer, but go in anchored on #1 with #2 as the graceful pivot.

04 Career-development mapping

Which projects compound toward which careers.

Project	Skills you'd build	Compounds toward
AI on the EDGE	Edge inference profiling, model partitioning, latency / energy Pareto, Jetson / CUDA	Edge-ML / MLSys / inference-optimization roles; hardware-aware ML research; PhD in efficient ML
Distillation SAM→YOLO	Knowledge distillation, quantization / pruning, real-time deployment, a shippable artifact	Model-compression / MLSys; strong portfolio + paper for PhD apps
Trajectory Prediction	Trajectory DNNs, simulation, LLM / VLM decision layer	Autonomous-systems / CV research; agentic AI
Agentic Streetscapes	Agent-loop design, bounded-latency tool use, eval	Applied agentic AI (fast-growing hiring area)
Adaptive Data Collection	Online learning theory, regret analysis, multi-distribution robustness	ML-theory / research / PhD theory track
3D Models	3D reconstruction, Gaussian splatting, real-time vision	Graphics-ML / 3D vision (new track)
Surgery (both)	Diffusion / world models OR agentic vision + medical domain	Medical-AI; generative video
Environment (×3)	Detection + Jetson / DeepStream deployment	Applied embedded CV (narrow)

The compounding read

#1 and #2 are the only two that deepen your existing "I build the silicon under the code" identity rather than starting a new track. #4 (agentic) is the best non-hardware bet because Palimpsest already gives you a head start and the market for agentic skills is rising fast.

05 Interview playbook · Wed 5/27, 3:00pm Zoom

Tomorrow. Lead with the hardware-systems line.

Self-contained. The general strategy + Seok-vs-Kostic decision frame live in the brief.

My background is the hardware side of edge inference — I can make models actually run fast on the device, not just train them.

That's his lab's whole thesis. Say it early.

Two concrete pitches

Lead with the first.

Pitch · 01 · Primary

Edge–cloud partitioning of an urban VLM / detector under a latency SLA on Jetson Orin

Explicitly extending the Distributed VLMs result (vision encoder on edge, language on cloud). Offer to profile where the latency really goes and find a better split point.

Pitch · 02 · Graceful pivot

Distillation-vs-partitioning tradeoff study

Distill SAM → YOLO to a real-time edge student model and quantify accuracy / latency / energy against the partitioning approach. This is the cleanest one-summer deliverable if he wants something self-contained.

Papers to name-drop — one specific thing each

Distributed VLMs (PerCom 2025): "The +33% throughput from running the vision encoder on edge is exactly the partitioning angle I want to push on."
Monocular 3D Tooltip Tracking: "You already use the SAM2 → YOLOv11 distillation pattern; I'd love to apply that to an edge latency budget."
SASS — Streetscape Application Services Stack: "The >10× edge-compute throughput result is the kind of systems win I care about."

Smart questions to ask him

Is the summer edge work closer to the Distributed VLMs partitioning line or to a distillation / optimization track?
Which device — Jetson AGX Orin? Is Coral / NCS in play, or Jetson-only in practice?
Is the target deliverable a workshop paper + public code release?
Given I'm finishing my first semester, is this a realistic summer for me, or would you rather I take E6692 first and start in the fall?
On registration / funding: would this be ENGI E4900 (zero-credit) or ELEN E6001/2, and is it paid, credit, or volunteer?

Honesty landmines — have an answer ready

Landmine · 01

First-semester acceptance "very rare" + no Kostic course

Acknowledge it; counter with the hardware-systems differentiator and your shipped projects.

Landmine · 02

Coral / NCS mentioned in cold email

His doc lists them, so you're not off-base; just note Jetson is the house platform.

Landmine · 03

Deliverables regime

Show you welcome the daily-log / weekly-report / final-paper cadence. You already ship public GitHub + portfolio sites; say so.

Landmine · 04

International / on-campus-only constraint

Summer must be a Columbia-registered role; maps cleanly to ENGI E4900 / ELEN E6001-2. Frame it as "I'm set up to register either way," not a complication.

30-second pre-call checklist

2:55 PM Wed

Zoom link arrived? (he sends just before — check inbox at ~2:50pm)
Have open: this doc, CV link, the 4 portfolio URLs (riscv-p6, palimpsest-nyc, geometry-of-alignment, pvzfpga)
One sentence ready on each of: RISC-V core, DE1-SoC, Palimpsest, Gemma distillation
No collision with Seok's Wed meeting (you kept 3:00–4:00pm clear for Kostic)

06 Sources

Citations & further reading.

Courses: E4040 Neural Networks & Deep Learning · aidl.ee.columbia.edu/documents/neuralnetworksanddeeplearning; E4750 Heterogeneous Computing · aidl.ee.columbia.edu/documents/sigproccommonmmulticore; E6691 Advanced Deep Learning · aidl.ee.columbia.edu/documents/advanceddeeplearning; E6692 Deep Learning on the Edge · aidl.ee.columbia.edu/documents/deep-learning-on-the-edge
Papers: Distributed VLMs (PerCom 2025) · ieeexplore.ieee.org/abstract/document/11038735; Adaptive Data Collection (ICML 2025) · icml.cc/virtual/2025/poster/44118; Data-Driven Traffic Simulation (CVPR 2024) · arxiv.org/abs/2408.00943; Language-Guided Traffic Simulation (CVPR 2026 Workshop) · see AIDL publications; Boundless · arxiv.org/abs/2409.03022; Constellation dataset · arxiv.org/abs/2404.16944; SASS · arxiv.org/abs/2411.19714; Urban Transportation Digital Twin (IEEE T-ITS) · arxiv.org/abs/2501.10396; Towards Suturing World Models · arxiv.org/abs/2503.12531; Monocular 3D Tooltip Tracking (MDPI Electronics 2025) · mdpi.com/2079-9282/14/10/2075
Lab & funding: AIDL site · aidl.ee.columbia.edu; Publications list · aidl.ee.columbia.edu/zkpublications; Lab GitHub · github.com/zk2172-columbia; COSMOS testbed · cosmos-lab.org; Google Scholar · scholar.google.com/citations?user=TlPI8yIAAAAJ
Companion documents: projects/columbia-summer-2026/kostic-research-projects.md — extracted from Kostic's login-gated opportunities Google Doc, updated 2026-05-21; meetings/briefs/2026-05-27-kostic-seok-interview-prep.md — general interview strategy

Kostic project deep-dive — all twelve subprojects, analyzed

A six-section read, ranked by fit against a real record.

Sections

30-second ranking

Who Kostic is, and how his lab actually works.

The four-course training ladder

Funding & testbeds

The deliverables regime — know this going in

Eleven subprojects, four tiers — scored against a real portfolio.

Standalone · primary target AI on the EDGE — Optimization of Inference for DL on Constrained GPU / IoT Devices ★

What it is

Kostic-background link

Fit vs. you — HIGH

Predicted popularity — HIGH

Career value

Environment umbrella · strong secondary Distillation for Low-Complexity AI Models ★

What it is

Kostic-background link

Fit vs. you — HIGH

Predicted popularity & career value

Smart Streetscapes Trajectory Tracking / Prediction

What it is

Kostic-background link

Fit · popularity · career

Smart Streetscapes Real-Time Agentic Models for Smart Streetscapes

What it is

Fit · popularity · career

Smart Streetscapes Adaptive Data Collection and Annotation

What it is

Kostic-background link

Fit · popularity · career

Smart Streetscapes 3D Models for Dynamic Streetscapes

What it is

Fit · popularity · career

Surgery Agentic Vision Models

What it is

Kostic-background link

Fit · popularity · career

Surgery Video Models in Robot-Assisted Surgery (World Models)

What it is

Fit · popularity · career

All eleven projects on five axes, side by side.

Where the crowd will cluster

Which projects compound toward which careers.

Tomorrow. Lead with the hardware-systems line.

Two concrete pitches

Papers to name-drop — one specific thing each

Smart questions to ask him

Honesty landmines — have an answer ready

30-second pre-call checklist

Citations & further reading.

Kostic project
deep-dive — all twelve subprojects, analyzed