Faizan Ahmed
Founder, Educator, and Engineer
Favorite Moments
Top 3 at OpenAI Hackathon
NYC Tour Meetup
Sajjaad and Farzy
NYC Meetup
NYC Office Meetup
Aidan Partnership
EdTech Conference Setup
NYC Office Meetup
Meeting Tariq (our first investor) for the first time
The Dispatch
from San FranciscoWork
Headstarter·Co-Founder & Head of ML
Working on proactive coding agents to let engineers work across multiple repos at once
Amazon·BI Engineer Intern
Developed and deployed an ML model trained on 100M+ customers, projected to increase profit by $100M+
Bloomberg·Global Data Analyst Intern
Automated equity research with Python & Bloomberg APIs
Research
Neural Geometry
Mechanistic interpretability is what I keep coming back to, because language models think in geometry. A cinematic tour of the shapes concepts take inside them, drawn from my own research.
Walk inside the transformer
I spend my time reverse engineering how transformers actually work. Step inside a single block with me, through its matrices, attention, and MLP, to see where my findings live.
Do Eval-Awareness Probes Detect Evaluation, or Just Style?
Linear probes are increasingly proposed to detect whether a model knows it is being evaluated. Across eight open models from 0.6B to 32B, I show they detect benchmark formatting, not evaluation: holding an item's content fixed and swapping only its costume inverts the probe, AUROC 0.90 to 0.0003 at 8B, and the inversion holds through 32B.
View on GitHubFixing What Isn't Broken: Represented Evidence vs. Action in Coding Agents
A mechanistic-interpretability study of how a code model represents pass/fail test-transcript evidence when it picks an action. A single layer carries whether the tests passed or failed, and that signal causally moves the edit-vs-abstain decision margin. Yet a strong action prior keeps the model from ever choosing to do nothing. The evidence the model represents is not the action it takes.
View on GitHubCausal Execution-State Representations in Code LLMs
Do language models actually use the program state they track internally? The running value of a variable is linearly decodable from their activations, but causal patching shows that trace is bypassed: answers are recomputed from the raw input tokens late in the network. Decodable is not the same as causally used; the gap holds across Qwen3, phi-4, and a 4B–32B dense ladder.
View on GitHubHold Your Fire: Disruption-Aware Failure Monitoring for Coding Agents
A local, CPU-only monitor that reads a coding agent's partial trajectory and predicts when it's drifting toward failure, about 13 steps early, while abstaining until it's confident, so it doesn't disrupt runs that would have succeeded. A 1.14 MB calibrated model outpredicts GPT-5.5 and Claude Opus 4.8 as failure judges at roughly 10,000× lower cost, cutting needless interruptions of healthy runs from 79% to 7%.
View on GitHubVANTAGE: Hidden Rewrite Views for Speculative Code-Edit Decoding
Accelerating copy-heavy code-edit decoding without changing the prompt the model sees. When an edit says "rename user to account", VANTAGE builds a hidden rewritten view of the reference code to propose speculative draft tokens that literal prompt-lookup decoding misses, while the target model still verifies every emitted token. Up to 1.64× faster on controlled workloads with identical greedy output.
View on GitHubAutonomous Coding Agent Benchmarking (Soon)
Building an end-to-end autonomous system where AI coding agents ideate, build, deploy, and self-test web apps, then automatically fix issues through browser agents.
A Deep Learning Approach for Covid-19 & Viral Pneumonia Screening
Presented an automated and efficient approach to screening respiratory conditions using X-Ray images, 30 citations.
View PaperCountries Visited
11Podcast
Building headstarter, early struggles, and advice for aspiring founders
Ask Me Anything
loading...