Columbia CS · reinforcement learning & NL2SQL research

Teaching machines
to learn from their own mistakes.

I'm Hassam Gani. I build reinforcement-learning and text-to-SQL agents at Columbia.

DAPLab · RL + NL2SQL researcherMorningsideMunch · founder & CEO2× paper submissions in progress
Explore researchRésuméGitHub
Avid neural-network trainer

Researcher & Builder.

Half of what I do is research. The other half is maintaining an app a thousand students open daily. Keep scrolling to learn more.

Hassam Gani
Who I am

The research I enjoy most is the kind that affects real users: I'm most interested in LLMs—from the tokenization to the transformer and all the way up to the RL, which is what I currently most work on. I split my time between models that teach themselves and a dining app that users frequently give my team & I feedback on how to improve. I think in four languages and debug in all of them.

Hassam
Based New York, NYSpeaks English · Spanish · Portuguese · ArabicCurrently two papers & a dining app
01 — Research

Agents that get
better on their own.

Two active lines at Columbia's Data, Agents & Processes Lab, plus earlier ML-security work, each aimed at a paper.

Mar 2026 — now

Building a self-improvement framework (targeting AAAI / ICLR) that turns an LLM's own wrong answers into synthetic training data through a challenger–solver–critic GRPO self-play loop. Reproduced R-Zero across 4 base models and 7 math benchmarks, then showed its reported +6.49 gain shrinks to +0.89 under a properly measured baseline.

+0.0
avg acc, Qwen3-8B
+0.0
OlympiadBench
0+
tracked H100 runs
veRLGRPOLoRAvLLMModal · H100
01.2 — Research

A graph the agent
can actually reason over.

May 2026 — now

Authoring an ICLR submission showing that reasoning over a graph-based semantic layer instead of raw table schemas makes LLM database agents more accurate on BIRD and LiveSQLBench, beating SQL-agent, Spider-Agent, and ReFoRCE baselines. Wrote 9K+ lines of Python: auto-generated PuppyGraph schemas over 58 Postgres databases, a Cypher ReAct agent with a matched SQL twin, and a GPT-5 eval harness over 1,350 tasks.

0×
graph-agent accuracy lift
0
Postgres DBs mapped
0
benchmark tasks
PuppyGraphCypherPostgreSQLGPT-5ReAct
01.3 — Research

Catching malware
hiding in ML models.

May — Aug 2025

Built cpyAnalyzer, a code analyzer that maps function and class-hierarchy dependencies to detect malicious object transformations in pickle-based Python and CPython models, contributing to policies that loaded 79.8% of benign models while rejecting 100% of malicious ones. Trained 10+ custom pickle-based classifiers for vulnerability testing alongside 7 PhD students.

0.0%
benign models loaded
0%
malicious rejected
$0K
Kluge Scholar grant
Python / CPythonDockerAST analysisHugging Face
1K+
02 — Ventures

MorningsideMunch

The dining app for Columbia & Barnard I founded and lead as CEO. 1,000+ users in under a month, totalling over 30% of dining-plan holders.

0+
App Store users
0%+
of plan holders
0
dining halls served
0→1
requests per refresh

Behind it: a Cloudflare-bypass scraping pipeline running headless Chromium inside Vercel lambdas behind a Redis distributed lock, feeding a multi-tier cache and a hand-rolled HTTP/2 APNs service, cutting per-refresh load from 20+ requests to 1 across every hall.

SwiftUIVercelCloudflare WorkersRedisAPNs · HTTP/2
Visit MorningsideMunch
02.5 — Beyond the lab

A wider track record.

Ambassador · Breakthrough VenturesNov 2025 — nowPart of a founder community spanning 40 founders and $50M raised.
Co-Founder · Stealth StartupAug — Nov 2025Built early computer-vision product foundations at a stealth startup.
Teaching Assistant, COMS 1404 · Columbia CSJan — May 2025Delivered an ML (NLP & CV) curriculum built with faculty and PhD students through workshops and mentorship.
Frontend Developer · Columbia SIPAApr — May 2025Designed and maintained course sites for the Executive MPA program.
Penetration Tester · CyberPro TechJan — Apr 2022Pentested 15+ systems and documented 200+ vulnerabilities with Burp Suite, Metasploit, and Wireshark across 20+ security projects.
03 — Projects

xPredict · SpaceXAI Hackathon Honorable Mention

A live prediction market built at the xAI Hackathon (Grok × X API track) that streams X posts and prices event outcomes in real time, a temperature-scaled softmax over log-odds evidence with exponential decay and EMA smoothing.

0
LightGBM models
0
engineered features
0K
lines shipped in 24h

Real ML under the hood: two LightGBM models trained on resolved-market ground truth with leakage-free TimeSeriesSplit / GroupKFold CV, correcting engine prices in logit space, served from a Dockerized FastAPI microservice with a versioned model registry.

LightGBMFastAPIX API v2GrokSupabase
Source ↗
Open Source

FaceGuard · FaceID for macOS

A lightweight facial-recognition security tool that trains on your face in 6 seconds without a GPU, averaging it into a persistent 128-column numpy embedding. On wake it scans for intruders and emails you the 3 most-confident non-matches plus a clip. Detects the owner within 2 feet 100% of the time, running 65% faster with a tuned top-k.

GitHub ↗OpenCV · numpy · Python
FaceGuard
SmartSnake
Deep RL

SmartSnake · it plays itself

A deep Q-learning agent in PyTorch, an 11→256→3 network reading hazards, heading, and food to pick relative moves, with experience replay, TD(0) targets, and ε-decay. Jumped 733% by game 100.

Play ▶GitHub ↗
RAG Agent

myHomer

A campus-aware RAG advisor over a per-user ChromaDB index of Columbia resources, grounded, cited answers with query rewriting, multi-turn memory, and guardrails; an agentic add-on auto-registers for classes the moment seats open.

RAG · ChromaDB · React
ML Security

cpyAnalyzer

An ML code analyzer that draws class-hierarchy dependency graphs to flag malicious object transformations in pickle-based models before they can hijack a system. Backed by a $5,000 Kluge Scholar grant.

GitHub ↗
04 — Contact

Let's build something
that learns.

Open to research collaborations, internships, and hard problems in ML & systems.

Email meLinkedInGitHubRésumé