Founders’ Adda

Detailed Summary

  • Moderator (Archana Jahagirdar) opened the session, explaining the “founder’s Adda” format: each founder should talk only about the product (no fundraising, no market‑size slides).
  • The audience was asked to keep the language accessible – jargon is fine for AI‑savvy listeners but a brief, simplified version should be ready for non‑technical participants.

2. Founder Presentations

2.1 Ravinder Kumar – Technod8.AI (formerly “Technodate AI”)

Key PointsDetails
Mission“Automate automation” – create an AI‑driven assistant that lets any company design, deploy and troubleshoot industrial robotics/automation without deep expertise.
Background15 years in robotics, previously with the world’s largest industrial‑robot manufacturer (achieved 100 % shop‑floor automation in 2010). Noted the paradox: high‑end robots exist, but most Indian factories cannot exploit them.
Product – Agent TKIA three‑module suite: 1) Conceptualisation (AI helps design the robotic solution), 2) Deployment (auto‑generates robot programmes, wiring diagrams, commissioning steps), 3) Troubleshooting (real‑time fault diagnosis & re‑programming).
Technical ApproachInitially considered building a foundational model but pivoted to customer‑driven experiments. After pilot deployments (including Fortune‑500 customers and a proof‑of‑concept with the Indian Air Force), the team concluded a bespoke foundational model is still required.
Demo HighlightsShowed a UI where the AI produces full system architecture, robot code, and step‑by‑step assembly instructions. Demonstrated a CNC‑style generation for aerospace engine fault handling: user enters an error code → AI visualises the 3‑D component and outputs a repair workflow.
Strategic PartnershipsCollaboration with Dr Sumit Chopra (AI “godfather” Yan Lee Kun) and academic talent from IITs.
Current StatusDemonstrations at Hall 14 (conference expo). Early field trials, but no commercial rollout yet.
Q&A HighlightsFoundational‑model question: Ravinder argued that generic chat‑LLMs cannot handle the “complex workflow” required for industrial automation; a custom application layer is essential.
Open‑source vs. proprietary models: Even if OpenAI builds a specialized model, it still needs a domain‑specific application stack and on‑premise deployment because industrial data is proprietary.

2.2 Vaibhav Vats Shukla – Quansys AI

Key PointsDetails
MissionBuild India‑first end‑to‑end voice‑infrastructure that can run a call‑center without any human‑in‑the‑loop.
Market ContextIndian call‑center industry is a $55 bn sector (≈ 2 % of India’s GDP). Existing platforms rely on outdated “agent API” models.
Product StackA layered architecture:
1) Data Engine – generates massive multilingual speech datasets (including Indic languages).
2) Model Layer – custom LLM + speech‑to‑text / text‑to‑speech tuned for Indian accents.
3) Orchestration Layer – handles call routing, context management, API hooks.
Key PartnershipsIntegrated with OpenAI for foundational model licensing; partnered with Paytm, CRED, PropBotX for early pilots.
Pricing ModelPer‑minute subscription – customers pay a usage‑based fee (e.g., ₹2 /minute).
ScalabilityCurrent concurrency ≈ 50 simultaneous calls; roadmap to increase as model optimisation progresses.
Cost‑Savings ArgumentFor large BPOs (e.g., SBI Insurance), automation can cut 90 % of call‑center labour cost, reducing price per minute from ~₹30 to ₹2‑3.
Q&A HighlightsIntegration with telecom numbers: The system can be attached to a PSTN line via a websocket handshake, enabling a fully automated inbound agent.
Foundational‑model scaling: Quansys built its own data‑engine (award‑winning from the Prime Minister’s office) to avoid reliance on public datasets that cause “serving‑break” failures.
Regulatory compliance: Data is processed on bare‑metal European servers (Hetzner) rather than public clouds, ensuring tighter control over privacy.

2.3 Pradyum Gupta – Papli Labs

Key PointsDetails
Problem StatementExisting map‑navigation services lack real‑time situational awareness (e.g., gate closures, fog, accidents).
Core TechnologyVision‑first data pipeline: dash‑cams, CCTV and other roadside cameras feed raw video to a central platform; AI extracts semantic events (open/closed gates, traffic density, weather conditions).
Data ScaleOperates on ~100 PB of video from ~8,000 fleet units (e.g., Delhi Transport Corporation buses).
Use‑Cases Demonstrated1) Dynamic billboard pricing for JC Decox – shift from flat‑rate to impression‑based pricing, boosting revenue by ~45 %.
2) Autonomous vehicle edge‑case mapping for MG‑Motors – real‑time hazard detection (fog, broken dividers).
3) Transit optimisation for DTC – route‑rationalisation algorithm that matches capacity to demand, increasing passenger load and revenue.
4) Live news generation – converting raw city‑wide video feeds into searchable news bites.
Business ModelB2B SaaS: Tile‑based pricing (e.g., 1.5 L per 25 km² tile per day).
ChallengesData acquisition consent – convincing drivers and municipal agencies to share video streams.
Privacy compliance (DPDP) – front‑camera faces and number plates are blurred; no raw footage is ever released.
Q&A HighlightsPrivacy compliance: All video is stored on bare‑metal European servers; faces/plates are anonymised before processing.
Incentive model for data providers: Papli doesn’t pay drivers; instead, the operational efficiency gains (e.g., DTC saved ≈ ₹800 cr of lost revenue) act as the business incentive.
Scalability: The platform processes petabytes in near‑real time using custom bare‑metal GPU clusters (CDAC‑supplied supercomputers).

2.4 Meenal Gupta – EGRIO Pi Solutions (Imagix AI)

Key PointsDetails
MissionProvide AI‑driven precision imaging & treatment‑planning for cancer radiotherapy.
Regulatory & ComplianceHIPAA, ISO 13485, CEDESCO (India’s medical‑device certification), multiple patents pending.
Problem Landscape~20 million new cancer cases annually in India; shortage of oncology experts leads to delayed treatment planning and poorer outcomes.
Product Flow1) CT/MRI upload to secure cloud DICOM viewer.
2) AI segmentation (organ‑and‑tumor contouring).
3) Radiotherapy plan generation – reduces manual contouring from 60‑90 min to 5‑15 min (≈ 80 % time saving).
Data & PerformanceTrained on 5 M images (≈ 30 % Indian data from remote North‑East regions). Accuracy 92‑99 % depending on case complexity.
Deployments & ImpactGujarat pilot: 1 M chest X‑rays, 4 000 TB‑positive detections, 6 early‑stage lung cancers identified.
Radiotherapy plans: > 1 000 completed; 50 k chest X‑rays screened with 2 700 TB flags.
RecognitionInvited to Microsoft to demo solution; praised by PM Narendra Modi.
Q&A HighlightsTrust & Human‑in‑the‑Loop: AI assists but final approval stays with radiologists; no replacement of clinicians.
Data residency: All data processed on‑premise or secure cloud within India, complying with healthcare regulations.

2.5 Vivek Gupta – Indus Labs AI

Key PointsDetails
MissionBuild a complete Indian‑language voice operating system (speech‑to‑text, text‑to‑speech, LLM, speech‑to‑speech) that can be used by any enterprise to create voice agents.
Why Indian‑centric?Dialectal variation every ~20 km; global providers (Google, 11 Labs, Azure) deliver only a generic Hindi model, leading to poor user experience.
Platform Architecture1) Infrastructure layer – own GPU servers and on‑premise data centres (low latency, sub‑500 ms).
2) API layer – modular services (STT, TTS, emotion‑aware STT, voice cloning).
3) Workflow builder – no‑code UI to design call flows, integrate with CRMs, trigger webhooks.
Key DifferentiatorsDialect mastery across Indian languages (including regional accents).
Emotion‑aware STT (recognises happiness, anger, etc.).
Cost advantage: ₹2 /min vs. ~₹8 /min for 11 Labs (≈ 70 % cheaper).
Latency: 300‑400 ms end‑to‑end, enabling natural conversation.
Partnerships & ReachTelecom partners Airtel, Jio for SIP integration; B2B customers (banks, NBFCs, FMCG). Global white‑label partners in Dubai, Germany; expanding to Arabic, German, French, Mandarin.
ScalabilityHandles 1 000 concurrent requests; can spin up from 0‑1 000 in ~10 min.
Q&A HighlightsData residency: All training data resides in India, satisfying sovereignty requirements.
Foundational‑model strategy: Started with public datasets, then built proprietary data‑engine to avoid “serving‑break” failures.
Cost‑impact: For typical BPO call‑center workloads, total operating cost can be cut by up to 70 %.

3. Cross‑Session Themes & Discussions

ThemeInsights / Observations
Localization & Indian ContextBoth Quansys AI and Indus Labs AI highlighted the necessity of building indigenous data pipelines and models to handle Indian dialects, regulatory rules, and data‑sovereignty concerns.
Infrastructure Choice (Bare‑Metal vs. Cloud)Multiple founders (Quansys, Papli, Indus) deliberately used bare‑metal or private‑cloud (European data‑centres, CDAC super‑computers) to maintain low latency, privacy, and cost control.
Foundational Model vs. Application LayerRavinder (Technod8) and Indus Labs argued that generic LLMs are insufficient for domain‑specific workflows; a custom application layer and domain‑specific fine‑tuning are essential.
Business Model – Usage‑Based PricingMost startups adopted pay‑per‑minute or per‑tile pricing, aligning costs directly with value delivered (call‑center minutes, mapping tiles, radiotherapy plans).
Regulatory & Privacy CompliancePapli Labs and Imagix AI both described privacy‑by‑design measures (blurring faces/plates, DPDP compliance, HIPAA/ISO certification).
Scalability & Real‑World DeploymentsDemonstrations ranged from pilot projects with Indian Air Force / Fortune‑500 (Technod8) to large‑scale city‑wide video ingestion (Papli), showing early traction but also highlighting the need for robust data pipelines.
Human‑in‑the‑LoopImagix AI, Quansys AI, and Indus Labs emphasized that AI augments, not replaces human operators, preserving trust and regulatory compliance.

4. Audience Q&A – Representative Questions & Answers

SpeakerQuestionAnswer / Discussion
Ravinder (Technod8)“If OpenAI builds a dedicated industrial model, will you still need your own application layer?”Yes – even a bespoke model must be wrapped in an application framework that handles data privacy, on‑premise deployment, and complex workflow orchestration.
Vaibhav (Quansys)“How do you handle scaling when the original public data caused service crashes?”Built a proprietary data‑engine (award‑winning), now generating synthetic multilingual data in‑house; also migrated to bare‑metal servers, avoiding public cloud throttling.
Pradyum (Papli)“How do you stay DPDP‑compliant with billions of video frames?”All video is processed on European bare‑metal servers, front‑camera faces and license plates are automatically blurred; raw footage never leaves the secure pipeline.
Meenal (Imagix AI)“How do clinicians trust the AI contours?”AI assists radiologists; final acceptance is always manual. The system’s 90 %+ accuracy has been validated on‑site with multiple hospitals.
Vivek (Indus Labs)“What is the latency for a real‑time voice conversation?”300‑400 ms end‑to‑end latency (sub‑500 ms), comparable to human turn‑taking, thanks to on‑premise GPU clusters.
General“How do you price for enterprises versus SMBs?”Most adopt usage‑based models: per‑minute for voice (₹2 /min), per‑tile for mapping (₹1.5 L per 25 km² per day), per‑plan for radiotherapy (fixed fee per treatment).

Key Takeaways

  • Product‑first mindset: Each founder focused on a concrete AI‑driven solution (automation, voice, mapping, oncology, voice OS) rather than fundraising or market sizing.
  • Local Data & Infrastructure: All startups emphasized the need for India‑centric datasets, bare‑metal/private‑cloud deployment, and data‑sovereignty to meet latency, cost, and regulatory demands.
  • Application Layer is Crucial: Even with powerful foundational models, domain‑specific orchestration, UI, and compliance layers are indispensable for industrial, healthcare, and telecom use‑cases.
  • Usage‑Based Pricing Aligns Value: Pay‑per‑minute, per‑tile, or per‑plan models allow startups to scale with demand while offering measurable ROI to enterprise customers.
  • Human‑in‑the‑Loop Remains Standard: Across sectors (radiotherapy, call‑center automation, industrial troubleshooting), AI is positioned as an assistive tool, preserving human oversight for trust and regulatory compliance.
  • Regulatory Compliance Integrated Early: DPDP, HIPAA, ISO 13485, and CEDESCO certifications were built into the product pipelines, not added later.
  • Scalable Architecture Demonstrated: From 100 PB video pipelines (Papli) to multi‑language voice APIs handling thousands of concurrent calls (Indus Labs), founders showcased real‑world scalability.
  • Strategic Partnerships Accelerate Adoption: Collaborations with OpenAI, telecom giants (Airtel, Jio), government bodies (Indian Air Force, PM’s office), and large enterprises (Paytm, Fortune‑500) provide validation and market entry.
  • Future Challenges: Building foundational models for niche domains, maintaining privacy at petabyte scale, and monetising large‑scale B2B deployments remain the next hurdles for these founders.

End of summary.

See Also: