AI Product Builder & Systems Engineer  |  Bengaluru, India
Portfolio  •  LinkedIn  •  GitHub  •  brafnan26@gmail.com

┌── Engineering Philosophy ──────────────────────────────────────────────────────────────┐
│  • High-concurrency systems, database internals & black-box isolation verification     │
│  • Deterministic multi-agent pipelines & autonomous explainable AI workflows            │
│  • Benchmarking & observability tooling for LLM inference (TTFT, latency percentiles)   │
│  • Local-first, edge architectures & tactile 3D visualization (Three.js)               │
└────────────────────────────────────────────────────────────────────────────────────────┘

High-performance verification engine that infers dependency graphs from client histories to detect isolation anomalies.

┌────────────────────────┐      ┌────────────────────────┐      ┌────────────────────────┐
│ Client-Observed Txns   │ ───► │ Dependency Graph       │ ───► │ Iterative Tarjan's SCC │ ───► Minimal Cycle Witness
│ (Concurrent SQL Runs)  │      │ (wr, ww, rw Edges)     │      │ & Shortest BFS Cycle   │      (G0, G1, G-single, G2)
└────────────────────────┘      └────────────────────────┘      └────────────────────────┘

• Stack: C++20, Python 3.12, PostgreSQL 17, Cloudflare Pages
• Mechanics: Analyzes concurrent client execution logs, infers per-key version orders, constructs direct read/write dependency graphs, and discovers isolation anomalies via an explicit heap-allocated DFS Tarjan's SCC and BFS cycle recovery.
• Performance: Processes ~142,000 txns/sec single-threaded with sub-millisecond anomaly detection on $10^4$ operations at $< 45\text{ MB}$ memory footprint.
• Live Tool: isocheck.pages.dev  |  Interactive Playground: isocheck.pages.dev/#/playground

Dedicated load-generation and capacity-analysis laboratory designed around the execution mechanics of LLM serving.

┌────────────────────────┐      ┌────────────────────────┐      ┌────────────────────────┐
│ Controlled Workload    │ ───► │ Streaming SSE Client   │ ───► │ High-Precision Clocks  │ ───► Capacity & SLO Report
│ (Ramp-up, Concurrency) │      │ (OpenAI-compatible)    │      │ (TTFT, ITL, Velocity)  │      (p50, p95, p99 Latency)
└────────────────────────┘      └────────────────────────┘      └────────────────────────┘

• Stack: Python 3.12, FastAPI, AsyncIO, Render, REST / SSE
• Mechanics: Sits external to model servers to generate synthetic concurrency while recording monotonic nanosecond timestamps (perf_counter). Computes Time to First Token (TTFT), Inter-Token Latency (ITL), token generation velocity, and SLO compliance curves without premature rounding.
• Live Console: inferload.onrender.com

Autonomous enterprise administrative agent turning unstructured customer inquiries into verified invoices.

Customer Message
       │
       ▼
┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│ Parser Agent │ ──► │ Pricer & DB  │ ──► │ Drafter Agent│ ──► │ QA Reviewer  │ ──► Structured Output
└──────────────┘     └──────────────┘     └──────────────┘     └──────────────┘

• Stack: React 19, Node.js, Express, Google Gemini API, Tailwind CSS, Vercel
• Mechanics: 4-stage sequential agent pipeline that extracts order intents from unstructured conversations, queries pricing rules, drafts quotes/invoices, and runs an automated QA review pass before persisting to the ledger.
• Live Demo: biz-pilot.vercel.app

The 6-second autonomous UPI & invoice reconciliation platform built for the ChatGPT CODEX Hackathon 2026.

┌────────────────────────┐      ┌────────────────────────┐      ┌────────────────────────┐
│ Raw UPI & Bank Feeds   │ ───► │ 7-Node Explainability  │ ───► │ AI Chartered           │ ───► Audit-Ready Ledgers
│ (Vague SMS / UTR data) │      │ Decision Tree Pipeline │      │ Accountant Agent       │      (6-Sec Auto Resolution)
└────────────────────────┘      └────────────────────────┘      └────────────────────────┘

• Stack: TypeScript, React, Vite, Tailwind CSS, LLM Decision Trees
• Mechanics: Ingests chaotic bank feeds (PhonePe, Google Pay, Paytm, HDFC) and matches them to open invoices using a deterministic 7-node decision tree (Amount → Name → Date → Narration → UTR → Duplicate → Resolution) paired with an AI accountant reasoning agent.

Distraction-free nature observation journal engineered for the Hacktoberfest 2026 DEV Challenge ("Touch Grass").

Field Notes / Photos ──► On-Device Local Storage ──► Edge / Local LLM ──► Ecological Synthesis
(Zero Telemetry)          (100% Client-Side)        (Llama 3.2 / Qwen 2.5)   (Structured Field Log)

• Stack: React, Node.js, Cloudflare Workers AI (@cf/meta/llama-3.2-3b-instruct), Ollama (Qwen 2.5 7B), Cloudflare Pages
• Mechanics: Local-first privacy architecture with on-device localStorage, zero user tracking, and edge AI synthesis that converts fragmented field notes into structured biological records.
• Live App: fieldnote-byg.pages.dev

Intelligent recovery agent developed for the Razorpay AI Buildathon.

• Stack: React 18, Vite, Node.js, SQLite, Anthropic Claude 3.5 Sonnet, Framer Motion
• Mechanics: Ingests gateway webhooks, isolates root transaction failure causes (auth timeout, balance, network drop), and calculates adaptive retry schedules alongside contextual customer outreach.

Real-time 1v1 brain training platform replacing doomscrolling with sharp, competitive mental duels.

• Stack: Expo SDK 57, TypeScript 5.x, Supabase Realtime, React Native
• Mechanics: Sub-50ms synchronized multiplayer logic and memory duels over Supabase Realtime channels with persistent ELO rating and anti-cheat state verification.

Civic grievance redressal portal integrating interactive Three.js 3D city ward visualization.

• Stack: React 18, Vite, Three.js, Tailwind CSS, React Router
• Mechanics: Classifies municipal defect reports, computes automated severity scores (0–100), and projects live ward-level status over an interactive Three.js 3D infrastructure model.

┌──────────────────┬────────────────────────────────────────────────────────────────────────┐
│ Domain           │ Technologies & Tooling                                                 │
├──────────────────┼────────────────────────────────────────────────────────────────────────┤
│ Languages        │ C++20, TypeScript, Python 3.12, JavaScript (ES6+), SQL                 │
│ AI & Agents      │ Multi-Agent DAGs, Google Gemini API, Claude 3.5 Sonnet, Ollama,        │
│                  │ Cloudflare Workers AI (LLaMA 3.2), LangChain, Decision Tree Engines    │
│ Systems & Perf   │ Concurrency Verification (SCC/BFS), LLM Load Testing (TTFT/ITL),      │
│                  │ Server-Sent Events (SSE), WebSockets, AsyncIO                          │
│ Frontend & 3D    │ React 18/19, Next.js, Vite, Three.js, React Native (Expo), Tailwind    │
│ Backend & Data   │ Node.js, Express, FastAPI, PostgreSQL 17, SQLite, Supabase, Redis      │
│ DevOps & Cloud   │ Cloudflare Workers/Pages, Docker, Render, Vercel, Prometheus, Git       │
└──────────────────┴────────────────────────────────────────────────────────────────────────┘