Open to work · graduating Feb 2027

Hi, I'm

Ahmad Tawil

I build AI systems — and measure whether they actually work.

Final-year Software Engineering student at Braude College. I build LLM applications, data pipelines and backend services end to end, and judge them by measurements rather than demos: evaluation suites, trap cases, cost tracking and honestly documented limitations.

See my projects CV GitHub LinkedIn

ahmad@portfolio: ~

❯ whoami

software engineer · AI systems · evals

❯ cat status.txt

● open to work

● B.Sc. Software Engineering — Feb 2027

❯ ls projects/

diffwarden/bureaucracy-navigator/eval-hybrid-rag/driftwatch/

❯ pytest -q

190 passed · across 4 repos

❯

Python
FastAPI
LangGraph
Claude
OpenAI
RAG
Qdrant
BM25
RAGAS
LoRA
Unsloth
vLLM
llama.cpp
Hugging Face
PostgreSQL
Airflow
SQLAlchemy
Docker
GitHub Actions
Vercel
Cloudflare
pytest
Python
FastAPI
LangGraph
Claude
OpenAI
RAG
Qdrant
BM25
RAGAS
LoRA
Unsloth
vLLM
llama.cpp
Hugging Face
PostgreSQL
Airflow
SQLAlchemy
Docker
GitHub Actions
Vercel
Cloudflare
pytest

0
false positives
DiffWarden on 28 review cases, including clean refactors and decoys — about $0.01 per review.
0% → 61%
correct refusals
Bureaucracy Navigator after fine-tuning a 1.7B Hebrew LLM — and 2× more correct answers.
17–26%
evidence retrieved
eval-hybrid-rag showed retrieval, not the LLM, was the real bottleneck in RAG.
$0.11
per full run
DriftWatch benchmarks 4 LLMs × 40 tasks × 3 repeats and catches a 15% drop within 2 runs.

Selected projects

Four end-to-end builds — each shipped with its own evaluation and measured results.

01Deployed · Sep 2026

DiffWarden

AI pull-request reviewer, running as a GitHub App

AI code reviewers are noisy: they comment on style, repeat themselves on every push and miss the lines that matter. DiffWarden only reports real bugs, security holes and performance problems — and stays quiet on correct code.

HMAC-verified GitHub webhooks on FastAPI; replies 202 in 0.13–0.22 s and reviews in the background.
LangGraph workflow with parallel batch review and a verifier pass; posts one review with inline comments and one-click “Commit suggestion” fixes.
Line-labelled diffs, JSON-schema outputs and content fingerprints, so comments land on the right line and are never repeated.

Python
FastAPI
LangGraph
Anthropic API
GitHub Apps
Vercel
GitHub Actions
pytest

Code Eval results Playground

Example inline review comment

precision: 100%precision
recall on 22 seeded bugs: 95–100%recall on 22 seeded bugs
per review, ~7 s: ~$0.01per review, ~7 s

02Research + product · Sep 2026

Bureaucracy Navigator

A fine-tuned Hebrew LLM that explains Israeli bureaucracy

Kol Zchut has the answers people need about unemployment, taxes or dismissal — but in long, formal Hebrew. Can a small open model, running on a laptop, answer in plain Hebrew, cite its sources and admit when it doesn't know?

Built a dataset from 287 Kol Zchut articles with a 12B teacher model (vLLM), then fine-tuned DictaLM 1.7B with LoRA (Unsloth + TRL) in ~20 minutes.
Telegram bot with RAG (multilingual-e5 + Qdrant), served locally as GGUF via llama.cpp — no personal document leaves the machine.
Summarises PDFs and phone photos of official letters with OCR, Hebrew RTL correction and ID-number masking.

LoRA
Unsloth
TRL
vLLM
llama.cpp
Qdrant
FastAPI
Tesseract OCR

Code Model Paper

פוטרתי מהעבודה. מה מגיע לי?

ייתכן שמגיעים לך דמי אבטלה1 ופיצויי פיטורים2. חשוב להירשם לשירות התעסוקה בהקדם1.כל זכות · דמי אבטלהכל זכות · פיצויי פיטורים

מכתב_ביטוח_לאומי.pdf

מה כתוב במכתב: …
מה לעשות: …
עד מתי: …
ת.ז. ●●●●●●●●● (הוסתר)

Example conversation (Hebrew)

correct refusals: 0% → 61%correct refusals
letter-format compliance: 28% → 100%letter-format compliance
correct answers, 36% faster: 2×correct answers, 36% faster

03Live explorer · Sep 2026

eval-hybrid-rag

Is hybrid retrieval worth it for RAG? A measured answer

Hybrid search is often recommended for RAG by default. Instead of eyeballing answers, I built the same pipeline three ways — dense, BM25 and hybrid (RRF) — and measured each on a multi-hop benchmark.

RAG over 609 news articles (17,653 chunks) with grounded, citation-enforced generation and JSON-schema validation.
50-question benchmark with 10 unanswerable traps, scored with RAGAS, refusal rates, latency and paired bootstrap confidence intervals.
Dockerised FastAPI service with rate limits and a free retrieval-only fallback, plus a live explorer of every answer.

Qdrant
BM25
RRF
OpenAI API
RAGAS
FastAPI
Docker
pytest

Live explorer Report Code

Context recall

dense0.525

traps refused: 8/10

sparse (BM25)0.550

traps refused: 9/10

hybrid (RRF)0.625

traps refused: 9/10

Real results · 50 MultiHop-RAG questions, k = 5

context recall with hybrid: 0.525 → 0.625context recall with hybrid
trap questions refused: 9 / 10trap questions refused
retrieval vs 10–12 s LLM: 38–83 msretrieval vs 10–12 s LLM

04Data engineering · Aug 2026

DriftWatch

A watchdog for silent LLM quality drift and cost

LLM providers update models behind the same name without any error — the API keeps returning 200 while quality shifts. DriftWatch treats evaluation as infrastructure that runs on a schedule and alerts when something changes.

Apache Airflow 3 pipeline (7-stage DAG, dynamic task mapping) that benchmarks OpenAI and Anthropic models on a versioned 40-task suite.
Deterministic and LLM-graded scoring, plus a rolling z-score drift detector with version-aware baselines and Slack alerts.
9-table PostgreSQL schema with cost caps, idempotent re-runs and a zero-cost backfill DAG.

Apache Airflow
PostgreSQL
SQLAlchemy
Alembic
asyncio
Docker Compose
Slack
pytest

Code

#alerts · drift on reasoning

How an alert fires (illustration)

results, 0 failures: 480 / 480results, 0 failures
false-positive rate: 3.2%false-positive rate
fewer reads after indexing: ~5×fewer reads after indexing

How I work

Measure, don't eyeball
Every project ships with an evaluation: a 28-case review benchmark, 232 held-out test examples, a 50-question RAG benchmark, simulated drift runs.
Test the traps
Decoys, unanswerable questions and pure-noise runs check that a system stays quiet when it should — not only that it finds things.
Report the honest result
Hybrid's gain was “suggestive, not conclusive”; the verifier pass removed nothing; 12% over-refusal is documented. Limitations are part of the result.
Keep cost and risk bounded
Per-run cost caps, rate limits with free fallbacks, PII masking, least-privilege permissions and secrets kept out of code.

Research & notes

My papers, plus notes where I take apart other people's papers and my own experiments.

Latest notes

Oct 2026When hybrid search loses: a lesson from Reciprocal Rank FusionMy experiments

Visit the R&D page

Skills & tools

AI & LLM

RAG
LLM evaluation (RAGAS, LLM-as-a-judge)
Fine-tuning (LoRA, Unsloth, TRL)
LangGraph
OpenAI & Anthropic APIs
vLLM
llama.cpp
Qdrant
Structured outputs

Backend

Python
FastAPI
asyncio / httpx
REST APIs
Webhooks
GitHub Apps
Telegram Bot API
Pydantic

Data

PostgreSQL
SQL
SQLAlchemy
Alembic
Apache Airflow
Query optimisation (EXPLAIN)
Hugging Face Datasets

DevOps

Docker
Docker Compose
GitHub Actions
Vercel
Cloudflare
Git
uv

Testing & QA

pytest (190 tests across these projects)
Evaluation harnesses
Integration tests with mock servers
Regression suites
Error analysis

Education

B.Sc. Software Engineering
Braude College of Engineering
Expected Feb 2027

06 · Get in touch

Let's build something that actually works.

I'm looking for a junior role in software, AI, backend, data or QA. Email is the fastest way to reach me.

ahmadtawil.se@gmail.com CV LinkedIn GitHub

Ahmad Tawil

Selected projects

DiffWarden

Bureaucracy Navigator

eval-hybrid-rag

DriftWatch

How I work

Measure, don't eyeball

Test the traps

Report the honest result

Keep cost and risk bounded

Research & notes

Skills & tools

Education

B.Sc. Software Engineering