Someone types
"I can't sleep again." Everything I build is about what happens in the next second.
Head of AI at Wysa·Bengaluru
I build AI that listens.
I lead AI at Wysa, building agentic systems for mental healthcare at scale: memory, voice, safety, and the infrastructure that keeps every reply fast and reliable.
Listening
Reading tone and context
0.98F1 on critical risk detection
89%fewer false positives in clinical escalation
12 → <1months to launch a new language
90%less compute from inference optimisation
Follow one message
01
02:04 · Local time
02
AI gateway
03
Safety and context
04
Reply
The gateway routes it
It picks the right model for the moment and keeps a fallback ready, so a stalled provider never reaches the person.
Risk is checked first
Safety comes first, while the system recalls what matters from earlier conversations.
The reply arrives
In the person's own language, checked by output guardrails, with a safe fallback ready if anything fails.
Selected work
BuiltLedTestedEvaluated
01 · MemoryArchitected and built
Conversational memory
Continuity across conversations: the assistant remembers what matters and brings it back at the right moment.
02 · OrchestrationBuilt
Agentic orchestration
Agents that work in parallel on safety, context and the reply, so every turn is careful and still fast.
03 · SafetyLed
Safety harness
Risk detection, guardrails, escalation and safe recovery, evaluated continuously against clinical risk categories.
04 · ConversationLed
Conversation design
Working with product, content and clinical teams on how AI conversations should feel: when to listen, when to reflect, when to guide.
05 · InferenceArchitected and built
AI gateway
One interface across model providers, with smart routing, fallbacks and reliability controls built in.
06 · ToolingBuilt
Prompt platform
A playground to write, test, version and ship prompts safely.
07 · ModelsEvaluated
Model strategy and self-hosting
Choosing the right model for each job across hosted and self-hosted options, balancing quality, latency, cost and privacy.
08 · VoiceExplored
Voice
Voice experiments and audio tooling, where latency and conversational behaviour are most exposed to the person.
09 · GovernancePrimary contact
AI compliance
Vendor assessments, privacy and data-retention requirements, and the primary contact for AI security compliance and audits.
Next: how I got here.
Keep scrolling
2016 → now
72 BPM · resting heart rate
The toolkit
What I reach for when a conversation has to be safe, fast and kind.
Agents · voice · AI safety · healthcare AI
Say hello.
akshatsurolia0@gmail.com