Abhisek Sahoo
Founder @ mayro.ai
Experience
- Full-timeJune–Oct 2025
Founding Engineer
- Designed event-driven distributed video processing pipeline with FFmpeg using serverless functions and message queues, supporting 10K+ concurrent media jobs with fault-tolerant retries
- Integrated LR-ASD-based active speaker detection into video processing workflows, improving speaker attribution accuracy for automated cropping, scene composition, and camera switching
- Orchestrated distributed GPU workloads on Modal, implemented memory snapshots, GPU snapshotting and dynamic batching to improve throughput and reduce inference cost per media-minute by ~15%
Full-timeMay–June 2025Software Engineering Intern
- Engineered forced alignment pipelines using Montreal Forced Aligner, for phoneme-level synchronization and frame-accurate editing workflows
- Engineered spatiotemporal face tracking pipelines with optimized bounding box smoothing, enabling adaptive framing and layout-aware composition
- Contract
Frontend Engineer
- Orchestrated asynchronous inventory workflows using TanStack Query mutations and Zustand-driven state machines, improving reliability of stock updates, purchase order transitions, retries, and side-effect handling.
- Optimized Next.js App Router performance across inventory dashboards and product catalog views through route-level code splitting and efficient asset loading, improving Core Web Vitals and reducing bundle size.
- Worked directly with ecommerce merchants to debug Shopify inventory sync discrepancies, troubleshoot purchase order failures, and resolve product-level stock mismatches across storefronts.
- InternMay 2024
Engineering Intern
- Re-Architected slope code with GitHub-integrated AI review workflows that enforced custom style guides, checked external links, and surfaced actionable inline suggestions on documentation pull requests.
- Developed automation workflows to monitor CLI definitions and OpenAPI specifications for changes, generating repository issues with diffs and suggested documentation updates.
Agent Runtime & Orchestration
- Built a dual-runtime agent architecture separating ephemeral streaming harnesses from durable workflow execution, letting interactive conversations and long-running campaigns share the same orchestration core.
- Engineered a request-scoped harness that dynamically assembled identity-bound tools, opt-in retrieval sources, and GenUI on every turn, with guardrails, eval loops, and OLAP-backed tracing in ClickHouse for quality monitoring.
- Implemented human-in-the-loop tool governance across streamed calls, approval/denial states, commit handling, replay-safe history, and UI gating for destructive actions.
- Unified interactive, scheduled, and autonomous execution behind one harness, context layer, approval model, and tool plane.
Context, Memory & Retrieval
- Architected a seven-layer GTM context engine spanning signal ingestion, ETL, knowledge graph construction, episodic/semantic memory, hybrid retrieval, intelligence synthesis, and activation.
- Deployed graph-augmented hybrid retrieval with semantic and lexical search, query expansion, reranking, and graph-context propagation, reaching ~180ms P50 and ~480ms P99 under production load.
- Layered multi-tier agent memory across episodic outcomes, semantic ontology facts, procedural playbooks, and temporal logs with infer-on-ingest or verbatim modes for query-time retrieval.
- Built cross-source entity resolution and graph deduplication across CRM, advertising, product analytics, and ingested content with canonical identities and merge policies.
Durable Data & Workflow Infrastructure
- Built durable, event-driven ingestion pipelines with step-memoized workers, async event chaining, per-step retries, and indexing polls, decoupling save latency from multi-minute graph construction.
- Extracted context infrastructure behind dependency-inverted service contracts for ingestion, polling, mapping, and retrieval, enabling replaceable backends independent of the graph-store SDK.
- Implemented workflow suspend/resume with typed resume schemas at approval boundaries, letting campaigns checkpoint, pause for review, and continue without HTTP request lifetimes.
Generative UI & Streaming Systems
- Shipped a streaming generative UI protocol emitting server-side JSON Patch operations through catalog-typed client components for incremental artifacts, deterministic replay, and conversation persistence.
- Designed a dual-fidelity conversation model with full message parts for rendering/citations and compact inference projections for model calls, using a CQRS-style presentation vs. model context boundary.
- Implemented zero-buffer streaming persistence committing user messages before execution and assistant messages from the streamed graph, preserving first-token responsiveness and multimodal history.
- Engineered tiered context compaction with tool-output caps, stubs for older results, and a 400K-character ceiling, achieving 85%+ history reduction across tool-heavy sessions without rendering regressions.
- Integrated a unified MCP capability plane with scoped, schema-stable tool sessions and per-tenant auth, replacing ad-hoc toolkit hydration with a consistent discovery/execution contract.
GitHub
+
Writing
- 05/11/25Building Auto Reframe: Remotion + Mediapipe with interpolation
- 05/03/26I've Seen the Light. It's Not PMF.
- 05/03/26Mathematical Framework for Finding a Premium Startup Name with Available Domains
- 05/03/26AI Video Translation: MFA, WSOLA, Visemes, and Evals
- 09/16/26Epistemic Humility: Nobody Owns the Truth