Pull for light
RRROSHAN REDDY
00Intro

Quant × AI engineer — Hyderabad, INThink more · Do more · Say less

Grit. Grind. Grow.

I'm Roshan Reddy — Computer Science Graduate, class of 2025. Experimenting with AI & LLMs, learning and building around markets, designing interfaces.

  • QARP · Momentum · Smallcap Alpha
  • Fine-tuned advisory LLMs
  • Local open-source agents
  • Interfaces
EST. 2025 — INDIA Enter 17.3850° N / 78.4867° E
01 / AboutThe operator's logslow read

Think more, do more, say less.

Motto

Who I am

BE Computer Science with a specialization in AI & ML, class of 2025, Hyderabad. About a year of shipping in the real world since. My days split between the market's open, the training loop, and the browser's dev console.

What I'm studying

Equity markets, daily — factor behaviour, walk-forward discipline, and the honest parts of drawdowns. On the AI side: LLM fine-tuning, local inference, and agents that do real work on open-source weights.

Off the clock

Cricket, Formula 1, basketball. Outdoor sport and video games in equal measure.

Physics is a feature

Drag the ball.
Throw it. Watch the seam.

The same integration loop that prices my strategies moves this ball. Grab it, flick it, let it bounce.

02 / CraftWhat I dofour pillars

Four things I build.
Slowly, properly, on purpose.

01

Systematic Quant

Strategies with a walk-forward spine. Factor selection, portfolio construction, and honest out-of-sample validation across NSE 500 and smallcap bands.

  • Backtesting
  • OOS validation
  • NSE 500
02

Interfaces

Design and engineering in a single pass. WebGL scenes, realtime motion, and interfaces built for flow — not for menus.

  • WebGL / Three.js
  • GSAP
  • Realtime
03

Applied AI & LLMs

Advisors, not chatbots. Fine-tuned models for financial, tax and legal advisory, running as agents on local open-source weights — private and cost-controlled.

  • Fine-tuning
  • Local models
  • RAG agents
04

Data Infrastructure

End-to-end market data pipelines with fallback layers. No single point of failure between the tick and the signal — reconciliation and graceful degradation built in.

  • ETL
  • Fallback layers
  • Resilience
03 / WorkThe booklive signal · self-audited

Three strategies.
Live signal. Self-audited.

01 · QARP MODEL

Quality at a Reasonable Price

● LIVE SIGNAL

Systematic blend of high quality scores & attractive valuation metrics.

OOS WALKFORWARD CAGR0%
SHARPE RATIO0x
Full2016–26 OOS2021–26 stitched
CAGR33.30%44.85%
Sharpe1.431.83
Sortino1.772.22
Calmar1.091.68
Max drawdown-30.56%-26.68%
Ann. volatility22.38%24.33%
vs Nifty 50 CAGR13.02%11.63%
SELF-AUDITSharpe 2.41 → 1.83 verified

NSE 500 universe · 122.9% / yr turnover

02 · MOMENTUM LEADERS

Momentum Leaders (Strategy A)

● LIVE SIGNAL

Classic composite momentum with Trend Gate & regime defense filters.

OOS WALKFORWARD CAGR0%
SHARPE RATIO0x
Full2007–26 · 18.6y OOS2016–26 · 10.25y
CAGR26.83%34.48%
Sharpe1.001.19
Sortino0.410.50
Calmar0.431.09
Max drawdown-62.22%-31.49%
Ann. volatility27.91%28.51%
Hit rate62.3%61.8%
SELF-AUDITCAGR 48.0% → 34.48% verified

Strict liquidity universe · trained on 2007–15, tested 2016–26 (never seen)

03 · SMALLCAP ALPHA

Smallcap Alpha Engine

● LIVE SIGNAL

High-conviction smallcap factor selection with strict liquidity bands.

OOS WALKFORWARD CAGR0%
SHARPE RATIO0x
Full2016–26 OOS2021–26 stitched
CAGR44.52%48.97%
Sharpe1.671.91
Sortino2.022.28
Calmar1.022.24
Max drawdown-43.79%-21.88%
Ann. volatility24.42%25.14%
vs Nifty 50 CAGR12.72%11.63%
SELF-AUDITCAGR 65.4% → 48.97% verified

₹500–20,000Cr MCap · ~170 names · 118.6% / yr turnover

Systems I've built

Agent Fleet AI / LOCAL

Multiple working AI agents running on local open-source models — financial advisory, tax advisory, and a lawyer LLM for legal advisory. Private by default, fine-tuned for the domain.

  • Local open-source
  • Domain agents
  • Private
Advisory LLMs AI / FINE-TUNED

Fine-tuned LLMs for financial, tax and legal advice — grounded answers over hallucinations, tuned on domain corpora instead of prompted into compliance.

  • Fine-tuning
  • LoRA
  • Domain corpora
Market Data Pipeline DATA / E2E

An end-to-end pipeline for market data with fallback layers. When a source blinks, the next one takes over — the signal never starves.

  • ETL
  • Fallback layers
  • NSE
04 / RoadmapThe thirty positionsverified, not vibes

Thirty positions.
One desk.

A research pipeline that doubles as a résumé — 30 projects across four tracks, each one a published question, a method, and a machine-verified number. The five stars ship their own working labs below.

A · Factors B · Derivatives C · Portfolio D · Research desk
The five stars 5 / 30

    Five stars · flagship labs

    ★ 5 SHIPPED

    Each flagship ships a working interface,
    not a slide.

    05 / AI LabForty agents · zero cloudollama · whisper · bge-m3

    Forty agents.
    Zero cloud.

    105 local agent workflows built and run on the desk machine — whittled to the forty that matter. Every LLM call runs on Ollama: no API keys, no cloud, no cost. The five demos below replay real runs captured locally — they need no model at runtime.

    0flagship agents
    7categories
    0API keys
    100%local inference

    Live from the machine · five demos

    REAL RUNS

    Replays, not screenshots —
    captured from actual local runs.

    The library — forty selected

    Everything runs on the desk.

      ▶ = live-run this session · ⚙ = needs a free key or local DB · UI = Streamlit app. All LLM layers are local Ollama.

      06 / GameThe arcadenet · hoops · race

      Play is just
      risk management.

      Three games, one physics engine. A net session, a free throw, and a race day — drag to aim, release to play, steer to survive. Same discipline as the desk, with a scoreboard.

      Deliveries0 / 10
      Wickets0
      Best—
      NSE CRICKET LAB / NET 04 PACE —

      Bowl your first delivery.

      drag to bowl · release to deliver
      NSE HOOPS LAB / FREE THROW 60S SESSION

      Line up, load the arc, release. 60 seconds on the clock.

      drag to aim · release to shoot · swish = 2
      NSE F1 LAB / RACE DAY TIME TRIAL

      Buddh International Circuit, India. Three laps, beat your ghost. SPACE to start.

      ← → steer · ↓ brake · SPACE pause · R reset · 3 laps
      07 / RunbookThe trailgrit → grind → grow

      The runbook.
      Every edge leaves a trail.

      1. 2022
        Round 01 · first lap

        First year in equity markets

        Learned to read a candlestick before a balance sheet. First year of drawdowns, discipline, and deciding the market was a system worth studying.

      2. MAY 2025
        Round 02 · graduation

        BE Computer Science — AI & ML

        Graduated with a specialization in AI & ML. Left campus with working agents and a data pipeline, not just a transcript.

      3. 2025
        Round 03 · the deep dive

        In-depth market research

        Market structure, flows, factor behaviour. Where the edge actually lives — and where the drawdowns hide.

      4. APR 2026
        Round 04 · the AI phase

        Experimenting with AI & LLMs

        Fine-tuned financial, tax and legal advisors on open-source weights — running locally, private by default.

      5. APR–JUN 2026
        Round 05 · the book

        Three investment strategies shipped

        QARP, Momentum Leaders, Smallcap Alpha. Walk-forward audited and self-verified — out-of-sample Sharpe between 1.19 and 1.91.

      6. JUL 2026
        Round 06 · certified

        NISM Research Analyst exam

        Passed the NISM Research Analyst exam. The paperwork finally caught up with the work.

      7. NOW P1
        Pole position

        Grinding the next edge

        Equity markets daily. New models, new interfaces, same discipline. Grit, grind, grow.

      08 / ScoreboardThe numbers hold upaudited, not adjectives

      The work, in six numbers.
      Fully audited.

      Live strategies 0
      Best OOS Sharpe 0x
      Best OOS CAGR 0%
      AI agents shipped 0
      Research roadmap 0
      NISM research analyst ✓ 2026

      The strategy numbers each carry a self-audit line in the work section above — the honest figure, not the flattering one. If you want the full walk-forward, just ask.

      09 / ContactWrite mealways open

      Always open.

      Open to freelance, collaborations, and hiring. If you have a hard problem — markets, models, or interfaces — say the word. Expect a calm, direct reply.

      mancharlaroshan@gmail.com→

      Hyderabad, India — IST
      market open to close