Research

Papers we've published and in flight.

We publish research across all three product verticals — autonomous agent frameworks (Koovis Workforce), Indian equity momentum (WealthPilot), AI video + Telugu cinema (Koovis Studios). We commit to publishing negative findings alongside positive — when our own approaches don't work, we say so with numbers.

In SubmissionKoovis Workforce

Autonomous Agent Frameworks: A Comparative Benchmark Across Founder Workflows

50 tasks × 8 frameworks, published negative findings included

Rajesh Kolachana (Koovis AI Pvt Ltd)

We benchmark Koovis Workforce against 7 other autonomous agent frameworks (CrewAI, LangGraph, AutoGen, OpenAI Agents SDK, Claude Agent SDK, Hermes Agent, Lindy) across 50 multi-domain tasks representative of solo-founder workflows — spanning coding, research, content, operations, and strategy. We evaluate on success rate, cost-per-successful-task, context retention, reversibility, and trust architecture. We publish negative findings about our own framework alongside positive ones. Full benchmark suite released as open-source.

Target venues:arXiv (cs.AI, cs.MA, cs.SE)SSRNNeurIPS 2026 Agents WorkshopAAMAS 2027ICLR 2027 Workshop on Agentic AI
In SubmissionWealthPilot

Do Regime Filters Add Value? An Ablation Study of Volatility-Adjusted Momentum in Indian Mid-Cap Equities

5-year OOS 2021-2025 · 2x2 factorial ablation · reversal of prior 4-year finding

Rajesh Kolachana (Koovis AI Pvt Ltd)

We run a 5-year out-of-sample (2021-2025) ablation study on Indian mid-cap momentum strategies with a complete 2x2 factorial design on volatility-adjustment and regime-gating. Neither enhancement adds alpha over the full 5-year window; the simplest variant (M1_NoRegimeNoVolAdj) is the only one passing all 8 validation gates (OOS Sharpe 0.96, alpha t-stat 2.05, DSR 0.63, 8/8). Finding reversed the 4-year OOS result, demonstrating OOS period selection's critical impact on ablation conclusions. We also publish a risk management ablation showing full 5-level RiskManager inverts alpha entirely.

Target venues:arXiv (q-fin.ST, q-fin.TR, stat.AP)SSRNPacific-Basin Finance JournalJournal of Empirical FinanceQuantitative Finance
In DraftKoovis Studios

Benchmarking AI Video Models for Feature-Length Pre-Visualization

Multi-provider evaluation with Indian-language focus

Rajesh Kolachana (Koovis AI Pvt Ltd)

We benchmark 5 AI video models (Runway Gen-4.5, Veo 3.1, Wan 2.2 14B, LTX-2.3 22B, Seedance 2.0) on full-length pre-visualization tasks spanning character consistency, motion quality, prompt adherence, and cost-per-second. First comparative study focused on Indian-language pre-viz (Telugu dialogue via Sarvam Bulbul + cultural adaptation). Published reproducibility artifacts (prompts, model versions, seeds, sample outputs).

Target venues:arXiv (cs.CV, cs.MM, cs.AI)ACM MM 2026SIGGRAPH Asia 2026CVPR 2027 workshop
ForthcomingKoovis Studios

Character Consistency in Long-Form AI Video

Multi-character pipeline design for feature-length pre-viz

Rajesh Kolachana (Koovis AI Pvt Ltd)

Best practices for maintaining character consistency across a full-length AI-generated rough cut. Asset-first strategy, cascading generation, reference-image patterns, multi-character shot handling with @tag prompt binding. Covers up to 3 simultaneous characters per shot.

Target venues:arXiv → CVPR 2027 workshop
ForthcomingKoovis Studios

Telugu AI Cinema: TTS and Cultural Adaptation

Multi-provider Indian-language TTS benchmark

Rajesh Kolachana (Koovis AI Pvt Ltd)

Comparative benchmark of Indian-language TTS providers (ElevenLabs Multilingual v2, Sarvam Bulbul v3, AI4Bharat IndicF5) on naturalness, emotional range, and cultural adaptation patterns for Telugu cinema pre-viz. Cost-per-minute analysis across hero and bulk dialogue.

Target venues:arXiv → IEEE Transactions on Multimedia

Get notified when we publish

We publish ~4-6 papers per year across all three verticals. Subscribe to be notified when each paper goes live on arXiv + SSRN.

Subscribe to research updates