IIT Madras-incubated Plenome deployed Ashwin AI at AIIMS New Delhi after winning BHASHINI's multilingual healthcare transcription challenge and Rs 35 lakh in pilot support, The Times of India reported. The system turns doctor-patient speech into structured records and prescriptions for high-volume government clinics. Founders say it handles Hindi, Tamil, Telugu, Gujarati and other Indian languages with about 10 to 12 second latency, and has generated more than 45,000 AI-assisted prescriptions, with Apollo Telehealth and NIC-supported hospitals also deploying it. That matters for public-hospital AI that must work in regional languages. The caveat is that accuracy and safety claims still need independent clinical audit.
Former Swiggy executive Anuj Rathi and former Zomato CTO Prashant Parashar launched Profound with $1.5 million from Stellaris, 3one4 Capital, and angels including Swiggy's founders, The Economic Times reported. Users build an AI representative through about a 30-minute voice conversation meant to capture career history and preferences for introductions and job matching. The team says it ran more than 1,000 pilot conversations and opened a waitlist in early August. That matters as India founders chase agentic products beyond chat wrappers. The caveat is that broader features are still rolling out and hiring outcomes are not independently audited.
Lightstorm secured a Rs 2,500 crore, 10-year fully underwritten debt facility from IndusInd Bank for AI-focused fibre among Indian data-centre hubs and onward to Singapore and Malaysia, The Economic Times reported. Proceeds back terrestrial corridors linking Mumbai, Chennai, and Hyderabad clusters plus the I-2SEA subsea system, where Lightstorm and Microsoft are co-owners. That matters as GPU campuses need low-jitter transport as much as chips. The caveat is that this is bank debt for connectivity, not equity into a model lab, so utilisation will decide whether the corridor becomes a true AI utility.
Blacksmith raised a $45 million Series B led by Peak XV at a $550 million valuation, nearly ten times its Series A mark, TechCrunch reported. GV and Y Combinator joined again, lifting total funding to about $58.5 million. The CI cloud helps teams build and test software before production and ships Codesmith, an agent that can fix failed checks as AI coding tools flood pipelines. Customers grew from roughly 700 to more than 5,000. That matters if validation becomes the bottleneck after generation gets cheap. The caveat is that GitHub Actions, cloud giants, and coding agents claim adjacent ground.
DeepSeek's API docs now list DeepSeek-V4-Pro-0813 as the live deepseek-v4-pro version, marking general availability after a long preview. The page keeps a 1 million token context window, up to 384K max output, thinking and non-thinking modes, and tool calling, at about $0.435 per million cache-miss input tokens and $0.87 per million output tokens. Company materials frame the cut as stronger on agent tasks such as terminal work and code engineering. That matters for teams that built agent stacks on cheap Pro pricing. The caveat is that many agent wins are vendor charts, and DeepSeek still warns of a future overall API price rise.
SpaceXAI released Grok 4.6 for multi-step agents, coding, knowledge work, and interactive visual projects. It says the model matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index and is live in Cursor, Grok Build, the SpaceXAI API, and partners such as OpenRouter, Vercel, and Cloudflare. Pricing starts at $2 per million input tokens and $6 per million output tokens, with a faster variant at twice that rate and double included usage in Cursor and Grok Build for week one. That matters as labs compete on agent endurance. The caveat is that several headline scores remain vendor-reported.
Researchers released SPIKE-Bench, an arXiv preprint accepted to COLM 2026 that tests whether language models emit biologically plausible, predicted high-risk protein sequences when prompted on toxin-design tasks. Across 32 models, compliance with risky prompts was common, a Functional Harmfulness Rate reached about 50.7%, and ordinary refusal rates poorly predicted functional risk. They also release BioSafe-Guard, a classifier meant to cut predicted risk while preserving benign biology help. That matters as biology copilots spread. The caveat is that predicted toxicity is not proof of a physical pathogen.
A Nature paper by Kasirzadeh and Gabriel, published Aug. 12, proposes agentic profiles that score AI agents on autonomy, efficacy, goal complexity, and generality. Each axis, they argue, raises different design and oversight questions, helping developers and policymakers separate narrow assistants from highly autonomous general-purpose systems. The taxonomy aims to make governance debates less one-size-fits-all as agents gain tools and longer horizons. That matters while regulators lack shared agent-risk vocabulary. The caveat is that a taxonomy alone creates no enforceable thresholds or audits.
SBI chairman CS Setty told FIBAC 2026 that AI-led banking should move past neat retail credit histories and into rural India, farmers, and small businesses, The Times of India reports. He said satellite imagery, digital records, and data-driven risk tools can widen farm lending if banks harden cyber defences, governance, and human oversight while leaving pilots behind. The speech adds a public bank voice after recent RBI remarks on board accountability. That matters for who gets scored into credit. The caveat is that a conference call is not yet a binding supervisory rule.
Sarvam AI and AI4Bharat released Indic DiarBench, an open Hugging Face dataset of about 108 hours of natural multi-speaker speech across all 22 scheduled Indian languages. It jointly scores ASR and speaker diarization with human-verified, speaker-attributed transcripts from meetings and in-the-wild audio, targeting overlaps and code-mixing that single-speaker tests miss. Sarvam says Indic-specialized pipelines lead commercial APIs on the joint metrics. That matters for tools that must know who said what. The caveat is that lab scores still need to hold on noisier telephony.
A parliamentary standing committee report, covered by Outlook Business, says MeitY spent only about 32% of IndiaAI Mission revised-estimate funds for 2025-26 by Dec. 31, 2025, and that Finance gave half of what MeitY sought for 2026-27. The Rs 10,371.92 crore mission also missed early fellowship targets even as officials stress GPU clusters and indigenous models. That matters for labs counting on subsidised compute. The caveat is that a lower budget estimate can reflect under-utilisation, not a permanent cut to the five-year outlay.
Shanghai hospital researchers published AbdomenNet in Nature Communications, a foundation model that reads non-contrast CT for 11 acute abdominal conditions plus risk-stratification tasks. Pre-trained on 103,989 exams and fine-tuned on 5,816 cases, it hit a macro AUROC of 0.919 on five emergent conditions across 2,528 external patients. AI assistance lifted radiologists' mean AUROC from 0.812 to 0.924 and cut median reading time by 52.5 seconds per case. That matters where contrast is delayed. The caveat is that the study is retrospective, so prospective use still must prove bedside safety.