// Real data pulled from Jwalin's GitHub READMEs. // Shared across all three design directions. window.PROFILE = { name: "Jwalin Shah", handle: "jwalin-shah", email: "jwalinshah13@gmail.com", linkedin: "linkedin.com/in/jwalin-shah", tagline: "AI Systems Engineer — Evaluation & Reliability · Tool-Augmented Reasoning · Privacy-First Systems", oneliner: "I build systems that measure, diagnose, and improve LLM behavior. Currently obsessed with grounded reasoning, deterministic computation, and on-device inference.", location: "San Francisco-ish", status: "open to research collabs", background: [ { org: "Sentient Arena", role: "Research Contributor (Cohort 0)", note: "Grounded financial reasoning, eval infra, failure-mode analysis" }, { org: "Skild AI", role: "Data Operations Lead", note: "Robotics data systems · 5 platforms · 25+ operators · task success +40%, overhead –50%" }, ], focus: [ "Grounded LLM reasoning", "Evaluation harnesses & telemetry", "Deterministic computation", "Tool-augmented agents", "Hallucination measurement", "On-device inference (MLX)", "Privacy-first architectures", "Robotics reliability", ], }; // 4 highlight projects + 2 supporting window.PROJECTS = [ { slug: "officeqa-arena", title: "OfficeQA Arena", kicker: "Grounded financial reasoning", blurb: "Multi-agent LLM system, top-tier (180+) on a 246-task financial benchmark. Built eval infra across 3,600+ tasks.", longBlurb: "Systematic exploration of grounded numerical QA over U.S. Treasury Bulletins. Demonstrated retrieval errors and tool selection — not model reasoning — drive the majority of failures.", metrics: [ { k: "score", v: "184.5", unit: "/246" }, { k: "pass rate", v: "75.0%" }, { k: "cost", v: "$1.71", unit: "for 246 tasks" }, { k: "iterations", v: "9", unit: "architectural generations" }, ], findings: [ "Simplicity wins — shell `grep` on raw TXT beat an 11GB SQLite + 10-component pipeline", "48% of failures = wrong table/row/column extraction", "Mentor framing > verifier framing (+13 points)", "Prompt ceiling sits around 70%", ], stack: ["MiniMax M2.5", "Goose", "OpenRouter", "Bash"], repo: "jwalin-shah/officeqa-arena", color: "amber", }, { slug: "openhuman", title: "OpenHuman", kicker: "Contributor · Personal AI super intelligence", blurb: "Open-source agentic desktop assistant. Local-first knowledge base, deep OS integration, one account for many providers.", longBlurb: "A clean desktop experience that goes from install to working agent in a few clicks. Workflow data stays on device, encrypted locally. Contributing to the upstream project at tinyhumansai/openhuman.", metrics: [ { k: "stars", v: "247", unit: "upstream" }, { k: "forks", v: "36" }, { k: "license", v: "GNU" }, { k: "stage", v: "early beta" }, ], findings: [ "Neocortex — local-first KB that compounds context across tools", "Subconscious — background self-learning loops", "Screen Intelligence — agent reads what's on your screen", "Inline autocomplete + voice (STT/TTS) anywhere on desktop", ], stack: ["Rust", "Local AI", "MCP", "Skills"], repo: "tinyhumansai/openhuman", color: "violet", }, { slug: "tensor-logic", title: "Tensor Logic → Cognition", kicker: "Working through Domingos (2025)", blurb: "A 3-scalar tensor-logic recurrence beats a 71M-param MLP at transitive closure — by 4+ orders of magnitude.", longBlurb: "A learning project, not a product. Each demo makes one idea click viscerally — from einsums to continual learning. Trained once on 16-node random DAGs; generalizes zero-shot to 1,500-node real Python import graphs.", metrics: [ { k: "TL params", v: "3", unit: "scalars" }, { k: "MLP params", v: "71M", unit: "fails at n=128" }, { k: "mean F1", v: "0.975", unit: "across 8 OSS pkgs" }, { k: "biggest graph", v: "1,532", unit: "nodes (sympy)" }, ], findings: [ "A logical rule and an einsum are the same operation", "Activation function picks the semantic (deduce / analogize / retrieve)", "TL is enormously parameter-efficient when a closed-form operator exists", "It can't magic one into existence — parity remains unlearnable", ], stack: ["PyTorch", "einsum", "EWC"], repo: "jwalin-shah/tensor-logic", color: "lime", }, { slug: "jarvis", title: "Jarvis", kicker: "Privacy-first iMessage assistant", blurb: "Local-first AI assistant on an 8GB M2 Air. MLX inference. <0.42s mean draft. <5% hallucination.", longBlurb: "End-to-end product system testing whether small local models can deliver practical reply drafting under strong latency and privacy constraints. Dual-path architecture: fast-path and background pipeline.", metrics: [ { k: "mean draft", v: "0.42s" }, { k: "p95 draft", v: "1.15s" }, { k: "retrieval Hit@5", v: "0.88" }, { k: "hallucination gate", v: "96.2%", unit: "pass" }, ], findings: [ "Local-first inference > cloud-first for privacy + predictable latency", "Template-first + generation fallback cuts cost and risk", "Tried-and-rejected: pure embedding classifier, polling updates, heavy fine-tuning", "Evaluated 37 model configs across latency / Hit@5 / hallucination", ], stack: ["MLX", "sqlite-vec", "BM25", "FastAPI"], repo: "jwalin-shah/jarvis-ai-assistant", color: "cyan", }, { slug: "inbox", title: "Inbox", kicker: "Unified terminal TUI", blurb: "iMessage, Gmail, Calendar, Drive, Notes, Reminders, GitHub — one TUI. MLX Whisper for ambient transcription.", metrics: [ { k: "sources", v: "7" }, { k: "stack", v: "MLX + Textual" }, { k: "model", v: "Qwen 3.5 0.8B" }, ], stack: ["FastAPI", "Textual", "MLX Whisper", "Qwen"], repo: "jwalin-shah/inbox", color: "rose", }, { slug: "robo-replan", title: "RoboReplan", kicker: "Tabletop planning RL env", blurb: "Custom RL env with FastAPI state transitions. GRPO training. ~80% success on tabletop robot planning.", metrics: [ { k: "success", v: "~80%" }, { k: "method", v: "GRPO" }, ], stack: ["FastAPI", "GRPO", "OpenEnv"], repo: "jwalin-shah/robo-replan", color: "orange", }, ]; // Stats — the "redesigned" widgets pull from these window.STATS = { totalRepos: 19, publicRepos: 19, // Real counts from `gh repo list jwalin-shah` (Apr 2026): 13 Python, 1 TypeScript, 2 unspecified // Shown as % share of primary language across public repos topLangs: [ { lang: "Python", pct: 81 }, { lang: "TypeScript", pct: 6 }, { lang: "Other", pct: 13 }, ], // Mock contribution heatmap — 7 rows (days) × 26 cols (weeks of last 6 months) // Higher density toward the right (recency) contributions: Array.from({ length: 7 }, (_, r) => Array.from({ length: 26 }, (_, c) => { const recency = c / 26; const noise = Math.sin(r * 13 + c * 7) * 0.5 + 0.5; const v = Math.max(0, Math.min(4, Math.round(recency * 3 + noise * 2 - 0.4))); return v; }) ), // Themes Jwalin cares about, with a "weight" themes: [ { label: "evaluation", weight: 1.0 }, { label: "grounded reasoning", weight: 0.95 }, { label: "MLX", weight: 0.9 }, { label: "privacy", weight: 0.85 }, { label: "agents", weight: 0.8 }, { label: "tensor logic", weight: 0.75 }, { label: "telemetry", weight: 0.7 }, { label: "RAG", weight: 0.65 }, { label: "RL", weight: 0.5 }, { label: "robotics", weight: 0.5 }, ], // Real recent pushes from `gh repo list` (Apr 2026) recentActivity: [ { date: "Apr 29", what: "tensor-logic — runnable demos for Domingos paper", repo: "tensor-logic" }, { date: "Apr 29", what: "fafsa-engine — neurosymbolic FAFSA SAI pipeline", repo: "fafsa-engine" }, { date: "Apr 27", what: "agihouse — ambient agent for Even Realities G2", repo: "agihouse" }, { date: "Apr 26", what: "quicksilver-zk — pure-Python ZK proof system", repo: "quicksilver-zk" }, { date: "Apr 24", what: "proof-of-action — provable privacy boundary (Ship to Prod)", repo: "proof-of-action" }, { date: "Apr 18", what: "officeqa-arena v5 — 184.5 / 246 best score", repo: "officeqa-arena" }, ], };