Sai Krishna Jasti
Senior AI/ML Engineer

AI systems built for reliability, scale, and decisioning.

I design and ship agentic and ML systems that are reliable, observable, and cost-aware - from research to production.

Currently:
_
Interested in applied AI systems that prioritize reliability and decision quality.
Highlights
Reactive allocation arbitration
Built a multi-agent allocation arbitrator that resolves post-scramble supply constraints with optimization + human-in-the-loop gates.
Real-time ML platforms
Delivered streaming and low-latency inference systems for decision support and personalization.
Evaluation & reliability
Developed backtesting, model ranking, and uncertainty controls to keep decisions stable under drift.
Systems

Selected Work

Principal AI Engineer I - Ascentt

Agentic Allocation Arbitrator

Agentic allocation arbitration for post-scramble supply conflicts with policy constraints, optimization, and HITL governance.

Problem Post-scramble allocation conflicts across regions and constraints.
Approach Multi-agent strategy synthesis + constraint translation into optimization parameters.
Outcome Deterministic allocations with explainable rationale and real-time traces.
AllocationOptimizationGovernanceDecisioning
Quick view
  • Inputs: demand, supply, policy constraints, and regional context.
  • Guardrails: HITL approval gates before execution.
  • Outputs: allocations + rationale + traceability.
Read brief
AI-Powered Debugging Platform for AI Agents

Epilog

Debugging platform for AI agents using traces, artifacts, and deterministic replay.

Problem Agent failures are hard to diagnose from logs alone.
Approach Trace + artifact capture with deterministic replay and AI-assisted RCA.
Outcome Faster debugging and more trustworthy agent behavior.
Agent ObservabilityDebuggingDeterministic Replay
Quick view
  • Trace capture + artifact comparison (UI, DOM, outputs).
  • Deterministic replay for failure analysis.
  • AI-assisted RCA suggestions.
Read briefGitHub
Real-Time Promo Cannibalization Detection

TrueLift AI

Streaming ML system that detects promo cannibalization and recommends actions in near real time.

Problem Promo cannibalization requires real-time detection and action.
Approach Streaming pipelines + low-latency inference + LLM explanations.
Outcome Near real-time decision support for retail ops.
Streaming AnalyticsDecisioningMLOps
Quick view
  • Minute-level aggregation + live signals.
  • Low-latency inference with explanation layer.
  • Action recommendations for retail ops.
Read briefGitHub
RESEARCH

Research in Progress

High-level summaries while publications are in progress.

AFDC - Adaptive Forecast Distribution Control

Risk-aware uncertainty control for forecasting under distribution shift.

Forecast ReliabilityUncertainty Control
Read brief
Regime-Aware Incremental Forecasting

Regime matching + checkpoint reuse to reduce retraining cost with stable accuracy.

Regime MatchingEfficiency
Read brief
EXPLORATIONS

Explorations

GeoSpot

AI location scout for entrepreneurs using public data + agent orchestration.

Coral ProtocolLocation IntelligencePublic Data
Read briefGitHub

Let's talk.

If you want to discuss or collaborate on AI systems, or just chat about tech, let's connect.

Contact