CROSSFIT
Heavy things, early mornings
● ML ENGINEEREDINBURGH, UKOPEN TO ROLES — NO SPONSORSHIP NEEDED
Bit-exact LLM reproductions. Autonomous agent systems. Fine-tuned SLMs on my own Grace Blackwell lab. Every claim on this site links to a repo, a chart, or a test that proves it.
max |Δlogits| = 0.0 — bit-exactarch axis: +0.1179 BPB [95% CI], 3 seedswithin 2.14x of Qwen3-0.6B val PPL at ~30,000x less data (n=1)dclm-edu: +0.7034 code BPB [95% CI], 3 seeds — largest effectpre-registered GRPO null — confirmed596,049,920 params reproduced
LIVE TELEMETRY — GITHUB + LEETCODE + HACKERRANK · ISR 6H
CONTRIBUTIONS1,649 contributions · past year
BUILDING THIS WEEK
LEETCODE
137/3991
solved
Easy89/954
Med.34/2084
Hard14/953
LANGUAGES
Python373 solved
C++30 solved
MySQL27 solved
Pandas15 solved
Java3 solved
Python1 solved
HACKERRANK · AI
#469
Artificial Intelligence · practice leaderboard
score 872.4 · hackerrank.com/yashbishnoi613
REPO
BuildFromScratch
★ 1 · pushed 2026-07-07 · Jupyter Notebook
REPO
max
★ 0 · pushed 2026-06-20 · TypeScript
FLAGSHIP · 01
BuildFromScratch
Single-file PyTorch reproductions of SmolLM2-135M and Qwen3-0.6B, both verified bit-exact against official HuggingFace weights (max |Δlogits| = 0.0) — then used as a lab bench: a three-build experiment applying 2026 papers at matched compute, with a proper statistical de-confound.
ATTRIBUTED EFFECTS — 3 SEEDS · ISO-FLOP · 95% CI · TEXT-LM-V2
PARITY GATE
0.0✓max |Δlogits| vs the HF reference. Bit-exact where measured error is exactly zero.
LARGEST EFFECT — DATA
+0.7034✓code BPB, dclm-edu vs FineWeb-Edu at matched tokens — ~2.3x the whole architecture bundle on the same corpus.
OPTIMIZER — EARLY TRAINING
+0.4743✓NorMuon vs AdamW, defended by an LR sweep 10x smaller than the gap. Scoped, not a scale claim.
MID-TRAIN ANNEAL — VERDICT: WIN
+0.2716✓vs an iso-token control that absorbs the LR-decay confound. The data, not the schedule, carries it.
THE RUN — HOVER TO SCRUB · FAITHFUL VS IMU-1 · TRAIN CE (n=1 CURVES)
Training-loss curve overlay. Move the pointer or drag on the chart to scrub. Left and Right arrow keys move the cursor one point; hold Shift to jump ten points; Home and End snap to the ends. Press Enter or Space to pin the tooltip and Escape to clear it. Clicking pins and unpins the tooltip.
THE DE-CONFOUND — TOGGLE THE AXES · P2 SUB-DRILL INSIDE THE ARCH BUNDLE
Deconfound Explorer
P1 AXIS3-SEED · 95% CI✓ VERIFIEDP2 SUB-DRILL3-SEED · 95% CI✓ VERIFIED
Toggle a training change on or off; the chart stacks its measured contribution into a predicted ΔBPB (the sum assumes additivity). Drill into the arch bundle to see how good that assumption really is.
DESCRIPTIVE — SINGLE RUN, IN-LOOP, MOTIVATION ONLY (n=1)
Three builds at 1.19B tokensDESCRIPTIVE · n=1✓ VERIFIEDin-loop val PPL (n=1, not suite-stamped)↓ lower is better
n=1 / DIRECTIONAL — 3 of 3 entries are single-run (no seeds); read as directional evidence, not an attributed effect
Faithful baseline (AdamW)28.65
Modernized (IMU-1 bundle)23.52■ best
Exploratory (partial-RoPE 0.25)29.54
select a row (Enter/Space or click) for per-seed detail
BuildFromScratch/Qwen3-0.6B/results/three_build.json @ fe18745556 · scripts/export_site_results.pyraw ↗
THE GAP — WHERE 1.19B TOKENS LANDS AGAINST THE ORIGINAL (LOG-X)
Gap to the original — val PPL vs training tokensDESCRIPTIVE · n=1✓ VERIFIED
hover, tap, or focus + ← → to inspect a point · Enter opens provenance
notes & provenance
metric: val PPL (held-out FineWeb-Edu, 204,800 tok, 50x4096 windows)
Single-run, in-loop comparison (n=1, descriptive). The ratio pairs are load-bearing and must never be conflated: 2.14x at ~30,000x less data (1.19B vs 36T tokens) and 3.5x at ~275,000x less data (131M vs 36T). Computed: 28.65/13.4 = 2.1381; 46.31/13.4 = 3.4560.
- Qwen3-0.6B/builds/2026-06-08_reproduce-faithful_qwen3-0.6b/results/original_vs_repro.txt
- Qwen3-0.6B/builds/2026-06-08_reproduce-faithful_qwen3-0.6b/results/qwen3_baseline2tpp_after.txt
- Qwen3-0.6B/builds/2026-06-08_reproduce-faithful_qwen3-0.6b/results/qwen3_after.txt
PARITY GATE — verify_run.py REPLAY · RECONSTRUCTED FROM COMMITTED verify.json
HONEST NULLS — REPORTED AT THE SAME STANDARD AS THE WINS
GRPO REWARD — 300 STEPS, NO LEARNING SIGNAL (n=1 TRACE)
GRPO reward, step by step
fraction verifier-correct per step · 300 steps
20 steps
SFT MASKING — THE WIN THAT WAS NOT (0.68 PPL → 0.009)
HONEST NULL✓ VERIFIEDΔPPL, SFT vs control (positive = SFT better)↑ higher is better
in-loop 'win' (confounded)0.6787
held-out, response-masked (corrected)0.0092
held-out, full-sequence (cross-check)-0.0064n.s.
select a row (Enter/Space or click) for per-seed detail
BuildFromScratch/Qwen3-0.6B/results/sft_null.json @ fe18745556 · scripts/export_site_results.pyraw ↗
GRPO AT 0.6B — PRE-REGISTERED NULL, CONFIRMED
HONEST NULL✓ VERIFIEDpass@1, gsm8k (n=100 items, T=0.8, 8 samples)↑ higher is better
n=1 / DIRECTIONAL — 4 of 4 entries are single-run (no seeds); read as directional evidence, not an attributed effect
SFT floor0.0112
GRPO0.01
random-reward gate0.01
RFT (iso-generation control)0.015
select a row (Enter/Space or click) for per-seed detail
BuildFromScratch/Qwen3-0.6B/results/grpo_null.json @ fe18745556 · scripts/export_site_results.pyraw ↗
CONTEXT EXTENSION — THE BASE-ARM DIP WE REPORT ANYWAY
Passkey ladder — retrieval accuracy vs context rung3-SEED · 95% CI✓ VERIFIED
hover, tap, or focus + ← → to inspect a rung
notes & provenance
Long-context passkey ladder (ecl-ladder-v1), rungs 512-8192, 5 depths x 8 keys = 40 probes per rung per cell; anneal arms pooled over 3 seeds (n=120/rung), base is a single checkpoint (n=40/rung). Wilson 95% CIs. The 4096 dip is BASE-ONLY (0.025 at its own trained length; both annealed arms hold 0.400 there) and unexplained — see paper §13. All three lines bump at 6144. Per-seed detail in ecl_ladder.json.
- Qwen3-0.6B/experiments/2026-06-30_qwen3-0.6b_midtrain-anneal/verdict.json
- Qwen3-0.6B/experiments/2026-06-30_qwen3-0.6b_midtrain-anneal/ecl_ladder.json
FLAGSHIP · 02 — A PROJECT YOU CAN TALK TO
Max — the assistant
Structured memory with hybrid retrieval, a self-written personality (SOUL.md), hourly proactive check-ins, and a fail-closed trust engine. The chat widget on this site isn't a plugin — it's the project itself.
INDEX OF WORKS — CLICK TO EXPAND · EACH OPENS A FULL CASE PAGE
Selected systems
Single-file PyTorch reproductions of SmolLM2-135M and Qwen3-0.6B, verified bit-exact against the official weights, then carried through a full research lifecycle — architecture, optimizer, data, post-training — where every cross-run claim is held to multi-seed CIs and iso-FLOP matching.
PyTorch · NorMuon · lm-evaluation-harness · iso-FLOP accounting
02Max
A personal AI assistant that accumulates context safely: structured memory with hybrid retrieval, a personality it writes for itself, hourly proactive check-ins, and a fail-closed trust engine. The chat widget on this site runs a version of it.
Bun · ONNX embeddings · Telegram · Slack · Claude / OpenAI / Gemini / Ollama
FIG 3.0 - ORCHESTRATION ENGINE
Six LangGraph orchestration patterns with a dual convergence gate (quality ≥ 8.0/10 AND factual accuracy ≥ 9.5/10), powering JobPulse: 15+ daily agents and a 4-gate recruiter-model pre-screen over 250 raw jobs/day at zero LLM cost.
LangGraph · GRPO · RLM · Python
04PRISM
FIG 4.0 - RECURSIVE RESEARCH
An agentic research system on the Recursive Language Model paradigm (MIT, Dec 2025): Claude as root model writes executable Python that orchestrates 10–50 local Qwen3-8B calls per query via vLLM, processing 1M+ token corpora with 100% step tracing.
Claude · Qwen3-8B / vLLM · Qdrant (BGE + SPLADE, RRF) · Langfuse · RAGAS
FIG 5.0 - PORTFOLIO FACTORY
Autonomous portfolio factory: job description in, published GitHub repo out. Ten orchestrated layers, 23 production patterns, Docker-sandboxed code generation, an independent zero-context reviewer agent, and a self-learning ReasoningBank.
Claude · DeerFlow patterns · Ruflo · Docker · policy gates
06Velox AI
FIG 6.0 - VOICE PLATFORM
Enterprise voice-agent platform handling real phone calls: Twilio + Deepgram STT + Deepgram/ElevenLabs TTS with a sub-2-second turn target on Gemini 2.5 Flash, a 5-layer anti-hallucination RAG stack, and multi-tenant Clerk auth with Stripe billing.
TypeScript · React 19 · Prisma · Gemini 2.5 Flash · GCP · Prometheus
FIG 7.0 - TALENT INTELLIGENCE
Kimi K2.6-powered recruiting agent: discovers LinkedIn candidates through your real logged-in Chrome via WebBridge, scores them on a 5-gate weighted rubric, drafts personalized outreach, and delivers a ranked shortlist to Notion + Telegram. Never auto-sends.
Bun · TypeScript · Kimi Agent SDK · SQLite · WebSocket dashboard
FIG 8.0 - FINE-TUNING BENCH
Reproducible LoRA/QLoRA fine-tuning pipelines: currently training Qwen3.5-9B on OpenThoughts-114k (bf16, r=32, alpha=64), with evaluation harnesses targeting HumanEval and GSM8K, tracked in W&B — the bench behind the @letsfinetune build-in-public series.
Unsloth · PEFT · TRL · bitsandbytes · Weights & Biases
On the bench: Foresight (predictive maintenance SaaS, built in 7 days) · code-graph-mcp (30 MCP tools, OXC + Louvain + Tarjan) · AgentForge Arena (agent tournaments, ELO leaderboard)
DWG NO. YB-LAB-01 · EQUIPMENT SCHEDULE
The lab — a datacenter on a desk.
FIG. 01 — DGX SPARK / GB104 MARKERS · HOVER OR TAB
RIG PHOTO — drops in here
COMPUTE
NVIDIA GB10 Grace Blackwell
~119 GB unified memory · aarch64 · DGX OS
SERVING
vLLM (PagedAttention)
runs PRISM's Qwen3-8B sub-LM at zero cost
MEMORY & TRACE
Qdrant · Langfuse · Postgres
hybrid vector search + 100% step tracing
GUARDS
safe_cuda.py · jax_safe_env.py
unified-memory caps for a shared box
SEE ALSO: /playground — RUN AN SLM IN YOUR BROWSER (WEBGPU)
Beyond the terminal → drag
Real photos land here soon — treated as full-bleed editorial spreads.
Horizontally scrollable gallery. Use Left and Right arrow keys to move one card, Home and End to jump.