Laxmi Sai Maneesh Reddy Jupalle

aqueduct runs bartowski's 17.8 GB Qwen3.8-27B Q4_K_M GGUF on a CPU-only machine with 8 to 16 GB of RAM. Layers that fit stay pinned; the rest stream from NVMe on every token. Greedy output is the same at every memory budget, thread count, and speculative k — no BLAS, no Python, no GPU.

• · Plan-before-allocate: from the GGUF header alone, pin whole layers from 0 until the budget fills, then stream the rest through a two-slot prefetch ring; refuse a budget the plan cannot meet
• · Hand-written AVX2 K-quant kernels (Q4_K / Q5_K / Q6_K) against Q8_K activations; bit-identical greedy ids at every thread count and at every memory cap
• · --spec 3 drafts with the checkpoint's own MTP head and verifies in one batched pass — 2.27× at 5 GiB, 1.78× at 11 GiB, no quality cost
• · Faster than llama.cpp's official AVX2 build at every rung on the same file; as shipped, llama.cpp cannot load this model in 5 or 8 GiB

• · Solo-built with Next.js 16, React 19, TypeScript, Supabase
• · 10+ workflows + Groq-powered Sparky AI for campus Q&A

Connects to Gmail, Google Calendar, and Slack, then opens to a synthesized "here is your day" briefing — what needs action, what's noise, in what order, and why. Opposite of cloud inbox tools: raw content never leaves the machine, and every briefed item carries a provenance handle to the exact source message or event.

• · Compile-time privacy: RawContent does not implement Serialize — raw payloads cannot enter outbound request bodies
• · Extraction via all-MiniLM-L6-v2 (ONNX/tract); synthesis via Qwen2.5-0.5B (candle GGUF) — both verified with the network physically disabled
• · Grounding enforced at four layers (memory, wire, disk, render); an ungrounded item is a hard pipeline failure, not a display glitch
• · Hybrid classifier (rules + embeddings); SQLite FK-grounded store; Tauri v2 shell consuming string DTOs only — never raw payloads

A keyboard-first workstation for the AI model market — mnemonic commands, multi-panel layouts, watchlists, two years of historical charts, and a unified entity model. Data from OpenRouter, LMArena, HuggingFace, arXiv, and lab feeds; normalized once on the server, served through a consistent REST + SSE API.

• · Entity resolver canonicalizes upstream naming (exact/alias/fuzzy tiers) and quarantines unresolvable rows instead of corrupting the entity table
• · Immutable snapshot history in SQLite (WAL) powers PX price charts and ARENA ELO/rank history — two-year backfill on first boot
• · Per-source pollers with independent cadence, backoff, and health — one dead feed goes stale in the status bar while everything else keeps ticking
• · 12 keyboard functions (TOP, DES, PX, ARENA, BENCH, WATCH, …) with ghost-text autocomplete and inline error bar

A single ~31 MB Rust binary that captures system loopback + microphone (no bot joining your calls), transcribes in real time through swappable STT backends, and exposes everything over REST/WebSocket. Bundled clients: an embedded web dashboard and an always-on-top copilot overlay that answers from the live transcript and whatever is on your screen. Privacy by default: with local Whisper, no audio leaves your machine.

• · Lock-free capture pipeline: WASAPI loopback + mic → Silero VAD → swappable STT (whisper-local, Deepgram, Groq, OpenAI-compat) → SQLite
• · Meeting notes: two-speaker live transcript, LLM auto-titles, summaries, full-text search, markdown export, optional synced audio playback
• · Copilot overlay (Tauri): hotkey captures active-window OCR + last 10 min of transcript and streams an SSE answer with context chips
• · No concealment, no continuous surveillance — screen capture only on explicit hotkey; localhost boundary defended against DNS rebinding

Ran a controlled fine-tuning study across three causal language models to answer one question: how much can fine-tuning alone improve a model, and what sets the ceiling? Built a complete QLoRA pipeline — base model → LoRA adapters → training → automated execution-based benchmarking → adapter merging → GGUF conversion → quantization → local deployment in Ollama.

• · Fine-tuned each model on the 75K-example Magicoder-OSS-Instruct dataset; measured coding ability with execution-based benchmarks against test cases
• · Qwen2.5-Coder-3B (modern, code-specialized): already saturated — fine-tuning barely moved the score, demonstrating the ceiling effect from the top
• · TinyLlama-1.1B (modern, general): 7/10 at baseline, modest clean gains
• · GPT-2-medium (2019, near-zero at code): 0/10 → 2/10, stayed at 2/10 even after 3× training — proving the plateau is a model-capacity limit, not a training limit
• · Adapted the pipeline to each architecture (GPT-2 Conv1D attention, missing chat template, 1024-token context); Drive-checkpointed training resilient to disconnects; solved BF16 vs FP16 and QLoRA-on-Conv1D issues
• · Key insight: pretraining sets the ceiling; fine-tuning helps you reach it — a modern 1.1B model out-scored a larger 2019 model with zero fine-tuning

Product layer on top of ByteDance's UI-TARS Desktop for safer, auditable, reproducible GUI agent execution — tamper-evident recording, human approval gates, dry-run simulation, and workflow export. Fully local; no data leaves the machine. Additive and opt-in; recording failures never break a live run.

• · Command safety gate + approval hook: evaluates run_command / run_script against destructive-action rules before execution
• · SHA-256 hash-chained Flight Recorder — append-only event log, content-addressed screenshots, integrity verification on read
• · Proof Pack export (shareable HTML evidence), Workflow Capsule, and Workflow Compiler with simulation-first guardrails
• · Simulation mode: agent observes and plans but skips real UI actions — dry-run workflows without touching the desktop
• · Team Workspace with roles, governance policies, and exportable audit reports
• · ~6.4k LOC across 9 commits; 19 new service tests passing; fully backward-compatible with existing agent control flow

• · Shipped v0.2 with background agent, system tray, egui settings GUI, and one-click installers (.exe, .dmg, .deb).
• · Rust engine: real-time notify watcher, SHA-256 dedup, rule-based organize, and nucleo fuzzy search across 100k+ files in <50ms.

• · Captures a page only after a real dwell threshold, then extracts the article body with Readability
• · Embeds on-device with a bundled all-MiniLM-L6-v2 (ONNX WASM); no model, font, or inference request ever leaves the browser
• · IndexedDB store (Dexie) with similarity edges and topic clusters: search by meaning, and related pages resurface while you browse
• · Manifest V3, Vitest in CI, JSON export and one-click delete of everything

• · YAML test files with assertions for JSON schema, contains, semantic similarity, LLM-judge, cost, and latency budgets
• · The same test runs against Anthropic, OpenAI, Gemini, and Ollama; local models are first-class so iteration is free
• · Self-hosted dashboard bound to 127.0.0.1 with run history, side-by-side compare, and trend charts

• · Six MCP tools: large files, cyclomatic complexity, dependency audit, dead-code detection, git health, and an aggregate 0–100 health report
• · Tree-sitter AST analysis for TypeScript, JavaScript, and Python; runs on demand, no background indexing
• · Works with Claude Desktop, Cursor, Windsurf, and Claude Code from one npx command; no databases or API keys

• · Ingests CSV, OFX, and QFX bank exports, dedupes them, and auto-categorizes with a hybrid rules + ML engine
• · 30-day cash-flow forecasts with confidence intervals; detectors for subscription creep, zombie subscriptions, spending drift, lifestyle inflation, and overdraft risk, with SHAP-backed explanations
• · Terminal vital-signs panel, dashboard, and PDF report; 533 tests, ruff and mypy --strict clean

Pre-arrival triage system for ambulances, built at the Microsoft-sponsored Hack with Chicago, where it took 1st place.

• · Multi-agent orchestration with LangChain and GPT-4: coordinated diagnostic, severity-scoring, and treatment agents for triage automation
• · Real-time streaming with Pathway, Kafka, FastAPI, and WebSockets for sub-second EMS data transmission; deployed on Azure with Docker

• · LLM document analysis with GPT-Nano and Gemini Flash; vector and graph search over Qdrant and Neo4j
• · Modular Extract, Query, Analyze, and Compare pipelines so decisions come from structured comparisons, not a single prompt

• · Multi-LLM pipeline for extractive and abstractive summarization, citation-aware retrieval, and contextual Q&A over uploaded papers
• · React, TypeScript, Node.js, and Supabase with JWT auth and row-level security; independent research project at UIC