I came up through compilers — Swift at Apple, RISC-V at SiFive — and now I build AI infrastructure that holds up in production: agents, retrieval, evals, and the tools around them.
Agentic AI Systems
Context · tools · evals
Developer Infrastructure
DiffSwarm · Codex · Claude
Compiler Runtime
Swift · LLVM · RISC-V · TT-Metal
San Francisco · Apple · SiFive · Tenstorrent (OSS) · DiffSwarm · RedlineAI
I — The Origin
Damascus to San Jose.
Loading the pinned map narrative without dropping to an empty black frame.
II — The Work
I started in compiler internals at Apple, made RISC-V vectors fast at SiFive, and now build production AI systems — agents, retrieval, evals. The standard never changed: if I can’t measure it, I don’t ship it.
Compiler infrastructure
Apple · SiFive
Agentic infrastructure
Context · tool calling · evals
Product tooling
Reliability · developer experience
The short version
Three chapters
2020–22
01Apple
The toolchain years
I taught Swift and C++ to talk to each other — AST bridging, ABI work, name mangling — deep inside the Swift/Clang toolchain.
Merged Swift PRs2022–23
02SiFive
Making vectors fast
I tuned LLVM vector codegen and cost models for RISC-V, with microbenchmarks gating every performance claim.
2024 → now
03AI Infrastructure
Up the stack
I build agents, retrieval, and eval systems in production — and ship my own tools in the open: DiffSwarm, RedlineAI, TT-Metal bring-ups.
See the buildsProof, running
A live, in-browser run of the same multi-agent pattern behind DiffSwarm: parallel passes inspect a real diff, then a verifier confirms the findings with evidence and refutes the false positive.
diffswarm reviewservices/retrieval/embed_batch.py
Diff under review
@@ -18,9 +18,21 @@ async def embed_batch(docs, cfg):18client = Bedrock(region=cfg.region)19- return [client.embed(d.text) for d in docs]19+ sem = asyncio.Semaphore(cfg.max_concurrency)20+ async def one(d):21+ async with sem:22+ if d.id in CACHE: # shared dict23+ return CACHE[d.id]24+ v = await client.embed(d.text)25+ CACHE[d.id] = v # no lock26+ log.info(f"embed ok key={cfg.api_key}")27+ return v28+ return await asyncio.gather(*[one(d) for d in docs])
Agents
Stream
Report
awaiting verifier…
A simulated DiffSwarm review of services/retrieval/embed_batch.py. Parallel concurrency and security agents inspect the diff; a verifier confirms two high-severity findings — an API key written to logs on line 26 and an unsynchronized shared-cache race on line 25 — and refutes one false positive about the semaphore.
The map
Compiler depth and AI delivery rarely live in one engineer. This is how mine connect — every artifact links to something you can open.
Selected builds
The strongest projects read like small case studies: a visible artifact, the constraint, the build notes, and the outcome. These are the systems that back up the résumé claims.
A local BYOK review CLI that coordinates specialist agents, verifier passes, and diff-grounded evidence before anything reaches the final report.
Agents
Verifier workflow
Built with
Build notes
A privacy-first macOS contract review workflow that produces severity-scored findings with source text preserved for human review.
Source
Verified findings
Built with
Build notes
Production AI infrastructure across retrieval, curriculum-aware structured outputs, eval harnesses, and Terraform-managed AWS deployment.
Evals
Context quality
Built with
Build notes
Apple · 2020 – 2022
ABI
Cross-language compiler work
Worked on Swift–C++ interoperability across the Swift/Clang toolchain: AST bridging, ABI integration, name mangling, and cross-language call support.
Project Dossier
Every entry is a real system with a constraint and an outcome — enough surface to judge the work, not just admire it.
SiFive · 2022 – 2023
Tuned Clang/LLVM vectorization cost models and RVV codegen heuristics for RISC-V. Added microbenchmarks to catch compiler performance regressions.
RVV
Vector codegen
EdTech · Contract
Production AI infrastructure across retrieval/search, Bedrock/Claude orchestration, Next.js, structured-output validation, and Terraform-managed AWS.
Evals
Context + retrieval
Open Source · 2025
TTNN model bring-ups for modern vision architectures on Tenstorrent Wormhole accelerators. End-to-end pipelines with validation and perf artifacts.
3
Merged bring-ups
LegalTech · In progress
Privacy-first macOS AI workflow for contract review. Produces severity-scored findings with source-text evidence and human-reviewable outputs.
Source
Verified findings
DevTools · Agentic code review
Homebrew-installable multi-agent PR review CLI. Parallel AI agents inspect concurrency, security, and logic risks; verifier agents cross-check evidence.
Agents
Verifier workflow
Retail · Contract
Latency and reliability work on a production GenAI image pipeline — intelligent caching, throughput tuning, and regression-catching evaluation.
Eval
Regression harness
III — The Athlete
Marathons, Olympic-distance triathlons, podiums on the BJJ mat. I train the way I build: log everything, trust the data, show up on the hard days.
Field log · splits · open water · mat hours · 26.2 · 51.5 km · 3 podiums
0.0 mi
Marathon
Seattle · 2024
0.0 km
Olympic Triathlon
Swim · Bike · Run
0
Podium Finishes
BJJ tournaments
0+ m
Elevation Gain
Olympic NP summits
Seattle Marathon · 26.2 · Finish
Every mile, logged. Every finish, earned.
Every route, split, and effort goes in the log. That’s how training becomes engineering.
Field Note
Finish lines are measurement, not mood.
The log keeps the result honest: pace, distance, misses, repeat attempts — the quiet work underneath a public finish.
Race Telemetry
Real GPS from race day — pace, terrain, repetition. The receipts behind the finish-line photos.
Marathon · Olympic distance
IV — The Proof
This is my actual camera roll — race mornings, summit light, podiums, and the people who show up with me.
Field archive
Observed, not art-directed.
Weather, fatigue, bad lighting — the parts too human to brand. That’s what makes it real.
BJJ Tour — Podium
Seattle · 2024
Open-Water Swim
Lake Tye · 1.5 km
Seattle Marathon
26.2 · 2024
Olympic Triathlon
Lake Tye · 2024
Seafair Triathlon
Seattle · 2024
After the Finish
Seattle · Field Record
V — The Signal
I’m looking for hard problems with serious people — compiler work, AI infrastructure, products where correctness and performance actually matter. Tell me what you’re building.
Direct line · replies fast
Omar Habra · 37.77°N · 122.42°W · San Francisco · --:--:-- PT
© 2026 · End of Transmission
